tldr-experts 0.3.1 → 0.5.0

package/CHANGELOG.md

@@ -1,5 +1,2328 @@
 # Changelog
 
+## 0.5.0 — 2026-09-02
+
+### Added
+
+- **The dashboard answers the five questions #85 left open (#93).** Each needed a decision
+  rather than a patch, and each was taken the CONSERVATIVE way: render what the files
+  already say, invent no interaction the page does not already have, and let nothing on a
+  read-only page run anything. Every choice below is a **maintainer call, subject to owner
+  review** — they are listed on the issue.
+  - **The Watchers tab draws the cards.** `05-watch/watchers/*.md` is read into the model as
+    the seven fields `Watcher` really declares — `id`, `epic`, `title`, `stories`, `repos`,
+    `status`, plus the optional `owner` (#70) — and the tab prints them instead of printing
+    the shape it wished it had. A `draft` card also carries **why**: the `absent:` sources it
+    cites under `## Signal`, which is the card's own rule for not being `verified`. The
+    reading is deliberately the SMALL one — the model resolves no `[src: …]`, calls no
+    `parseWatcherCard` and computes no `CardChecklist`, because that re-checks every citation
+    against today's working tree and would make a read-only dashboard the only screen in the
+    product that runs something. `tldrx watch check` stays the thing that checks, and the
+    page says so. A `verified` stamp over an `absent:` Signal is shown as what it is rather
+    than silently corrected: `watch check` re-stamps cards, and a viewer that disagreed with
+    the file would be a third opinion.
+  - **A `draft` card raises no attention card.** The page's rule — an alert means a run is
+    waiting on a PERSON right now, derived once in `waiting.ts` — is unchanged. An
+    uninstrumented signal is a fact about coverage that stays true until somebody
+    instruments it, and it belongs in a panel the way `budget.blocked` does.
+  - **Preflight refusals leave their rows on the page.** `04-build/preflight.yml` is read the
+    same way `budget.yml` is: read-only, additive, and through the reader that never throws.
+    A new **Base gates** section names each of the workspace's own gate commands, the repo
+    and base it ran on, its exit code and its `ok`/`failed`/`unmeasured` status — so a Build
+    that refused to start is no longer a stage that went backwards for no visible reason.
+    The alternative on the issue was to emit an EVENT instead; that is the bigger, better
+    change and it is not this one, because the owner's #85 decision covered the two files
+    that already existed and an event is a new write on a refusal path. Reading the file it
+    already writes is the same doctrine one step further. Red rows are drawn as rows, not as
+    alerts, for the same reason `budget.blocked` is.
+  - **`keep_worktrees` is one line on the run detail**, and only when it is set (#16). The
+    key is written only when true, so drawing `false` would put a row on every run in the
+    workspace saying what all of them do. `RunDocument` projects it tolerantly: anything that
+    is not the boolean `true` is not a promise to keep the worktrees.
+  - **A cancelled run says who closed it, when, and why.** All three facts already travelled
+    on `waiting.message` (#86); the run detail printed the status chip and dropped them. It
+    is one `kv` row carrying that sentence — not a second derivation, so the page cannot word
+    it differently from `tldrx run status`.
+  - **An annotated run wears a marker in the runs list** — one ✎ glyph with the note count in
+    its `title`, nothing more. #85 §1 asked for it and the wave declined to invent it. This
+    is the smallest thing that is true: no column, no badge count, no new sort key, and the
+    notes themselves stay on the run detail where they were. Explicitly **provisional** — a
+    count in a row is a design decision, and the first person with a real opinion about that
+    list should replace it.
+  - **`DASHBOARD_MODEL_VERSION` stays at 3.** `watch`, `preflight` and `keepWorktrees` are
+    three additions and nothing was removed; no existing field reads differently than it did
+    at v3. The argument for bumping is that `docs/dashboard-model.md` promised, under *What
+    is NOT in it*, that two named files were unread, and that promise is now void — but a
+    documented absence is not a field, and the number is for fields.
+
+- **The dashboard reads `budget.yml` and `events.jsonl` (#85).** The audit that filed the issue
+  found five gaps with one root: `buildModel()` read `run.yml`, the phase artefacts, the Plan
+  artefacts and the expert files, and nothing else — so five facts a reader went looking for were
+  nowhere on the page, each because it lives only in one of those two files. The owner's decision
+  (2026-09-02) was that the model may read both, read-only and additive. All five are now drawn.
+  - **Operator notes** (`tldrx note`, #46) get their own section on the run detail, with the actor,
+    the stage or phase, and the time. **All of them**, not the last three: `tldrx run status` caps
+    at three because a terminal has a bottom, and a run detail page does not.
+  - **Free review retries** (#78/#79) and the **attempt** each story is on. Both are event-only —
+    the story file carries `status` and `evidence` and no counters — so a story that burned both
+    attempts and was granted two free re-prompts read on disk exactly like one nobody had touched.
+    The plan table gains an `attempts` column (`1 of 2`, against the model's new `maxAttempts`), and
+    a **Reopens & retries** section shows the `story.reopened` arcs with their `fix`/`attempts`
+    reason and the operator's note. A reopen written before the `reason` key existed reads as
+    `attempts`, which is the only kind that existed, rather than as a blank.
+  - **`budget.blocked`** occurrences are listed with the phase, both economies' numbers, and — for a
+    dollar refusal — the exact `tldrx budget raise` command, short-by rounding included, pinned by a
+    test against `raiseCommand` because the renderer is serialised into the page and cannot import
+    it. Deliberately **not** an attention card: the page's rule is that an alert means a run is
+    waiting on a person NOW, and a refusal in the log is not evidence of that.
+  - **The `$0.00 of $25.00` progress bar on a host run is fixed properly.** A run whose `budget.yml`
+    says `economy: host-tokens` is not priced in dollars at all, so `ceiling_usd` governs nothing
+    and the bar stated a fraction of a denominator that does not apply. Such a run now reads in
+    TOKENS on both screens — the runs list and the run detail share one `dashSpendText`, because
+    suppressing the bar alone would leave the words `$0.00 of $25.00` making the claim the bar
+    was — metered against `ceiling_host_tokens`, the ceiling those tokens really are judged
+    against, which exists in no other file. A **Budget** panel carries the per-phase ceilings, `on_exceed`,
+    `warn_at_pct`, `on_host_tokens_exceed` and the per-phase economy. An unset phase economy is
+    reported as *inherits*, never as a choice somebody made.
+  - **`DASHBOARD_MODEL_VERSION` stays at 3**, and the issue asked. Eight fields were added and none
+    removed. `spentUsd` was the one with a case to answer — a consumer reading it alone is
+    demonstrably wrong about a host-attended run now that the ceiling can sit beside it — but it is
+    computed from the same `run.yml` key, holds the same number, and has meant "METERED dollars, a
+    lower bound when `unmeteredTasks > 0`" since v3 put `unmeteredTasks` and `hostTokens` next to it.
+    A field that gained neighbours did not change meaning.
+  - **Neither file is opened by the model**, and the ledger is walked **once per run**. `loadRunResult`
+    already parsed both for every run and this file had been discarding them, so the page costs the
+    reads it always did; the per-story facts come out of one pass rather than from
+    `readReviewLedger`, which re-reads the whole ledger per story and would have made a forty-story
+    plan forty passes over a file already in memory.
+  - **Absence and damage stay graceful.** No `budget.yml` and no `events.jsonl` renders exactly as
+    before, with none of the new sections. An unparseable `budget.yml` costs the panel and nothing
+    else. A torn ledger line costs that line and **says so** on the page, because "no operator notes"
+    over a damaged ledger is the same lie by omission an unlisted corrupt `run.yml` was.
+  - Fixed on the way: an event with **no `payload` key at all** parses fine through the tolerant
+    `EventLog.readAll`, and reading `payload.story` off it threw a `TypeError` out of `buildModel`
+    and killed the live server for the whole workspace. `TldrxEvent.payload` is typed non-optional;
+    the type is a claim about `validateEvent`'s output, not about what is in the file.
+
+- **The `tldrx drive` mandate carries its own preflight (#84).** Launching a cold unattended
+  session took SIX hand-run commands before the mandate could be pasted at all — `tldrx run
+  attend host`, then `tldrx run gates set` five times. Every one of them is a precondition of
+  the discipline the mandate exists to transfer, so an owner typing them by hand was doing the
+  driver's job for it, and a mandate that assumes its own preconditions only works where
+  somebody has already been careful. Both modes now open with a **`## Before anything: the
+  preflight`** section, ahead of the roles, that establishes attendedness (`tldrx run status
+  <run>`), the gate policy (`--json`, stage by stage) and `budget.yml` — and makes the driver
+  **state the ceiling it will honour** before the first turn. Any one it cannot establish is a
+  **refusal to start** that names the command that failed, because a driver that starts anyway
+  has spent money on a run whose gates it may not close.
+  - The two modes differ here exactly as they differ at the gate. `--unattended` may SET what
+    it finds wrong: `tldrx run attend host <run>`, and `tldrx run gates set <stage>:agent
+    --note "…"` for a stage the owner delegated, over a note **quoting the owner's own
+    delegation from the launch message** — so the policy change is signed by the owner's words
+    rather than the driver's judgement. `--attended` gets the mirror and may not: gates stay
+    human there, so it checks attendedness and the budget, reports a stage that is not `human`
+    where it was expected, and moves nothing. A test asserts the attended text contains no
+    `gates set <stage>:agent` at all.
+  - The `<run>` substitution (#75) covers the new lines like every other, `<stage>` untouched:
+    the pinned occurrence count moves from 5/7 to **7 attended / 10 unattended**.
+  - Both mandates stay inside `MANDATE_MAX_LINES` (119 unattended, 116 attended), paid for out
+    of the header, the driving block and the unattended gate. The sections that make the
+    document worth pasting — the three-role spine, evidence discipline, parking, review
+    calibration by stakes and budget honesty — are untouched.
+
+- **The retro aggregate has consumers: `--json`, the reviewer prompt, and a workspace taxonomy
+  (#74).** `tldrx retro --all` (#64) produced exactly the dataset that issue asked for — finding
+  class × count × runs × one cited example — and nothing read it. The reader existed; the loop it
+  was meant to close was open at the far end. Three seams close it.
+  - **`tldrx retro --all --json`** — a stable machine shape: `{version, root, runs, contributed,
+    deduped, classes, trends[{cls, count, runs, example{run, kind, text, src}}], findings[]}`. A
+    deliberate projection rather than a dump of the internal type, with its key sets asserted
+    literally by a test, so adding a field is a visible act and renaming one bumps `version`.
+    `--json` belongs to `--all` alone: closing one run WRITES a file and has nothing to parse, so
+    `retro --json` on its own is a refusal rather than a stringified sentence.
+  - **The adversarial reviewer is fed the workspace's top three classes** before it reads the
+    diff, so a review starts from what this team keeps getting wrong instead of rediscovering
+    `test-cannot-fail` on its own, run after run. Computed by the same `mineAll` over the same
+    workspace, so what the reviewer is told is exactly the top of what `tldrx retro --all` prints.
+    Additive and absent-safe in every direction — no runs, no findings, nothing but `other`, or a
+    broken taxonomy file, and there is no section at all, not a heading and not a blank line.
+    `other` is never offered: it names no defect to look for. Mined once per `tldrx next`
+    invocation, so a wave of six stories pays for it once and every reviewer in that invocation
+    gets the same prior. The section says out loud that it is a PRIOR and not a checklist — a
+    reviewer handed three defect classes and no framing finds three defect classes.
+  - **`.tldrx/memory/finding-classes.yml`** adds classes for defects the seven built-in ones do
+    not name (`version: 1`, `classes:` of `{name, rules}`, 1–16 each, rules compiled
+    case-insensitively). Extensions are tried AFTER every built-in rule and before `other`, so a
+    workspace class can only ever claim a finding the built-ins left unclassified — which is what
+    keeps an unbounded taxonomy testable, and leaves every shipped fixture immune to whatever a
+    workspace writes. A file that will not load is a REFUSAL naming the file, the class and the
+    rule — including a rule such as `.*` that matches every text and would swallow the taxonomy —
+    never a silent fallback, because a rule its author believes is running and is not would make
+    every count a lie. The Build reviewer never fails on it: a refusal costs the prior and prints
+    one line, so no story loses an attempt to a YAML typo.
+
+- **`CONTRIBUTING.md`, and a section an outside contributor can build a model provider from
+  (#27 companion).** #27 (a generic model-provider layer) was closed as parked, not rejected:
+  `TLDRX_CLAUDE_BIN` covers today's needs and a provider layer with no second provider is
+  speculation. So the seam is written down instead. The new file covers the contribution loop,
+  the four gates and what CI actually runs, the red-first rules (including "a test that cannot
+  fail is worse than no test", with the shipped example of one that could), and a
+  **Contributing a model-provider config** section: what `TLDRX_CLAUDE_BIN` does and does not
+  buy, `buildClaudeArgs`' exact command surface, the `stream-json` transcript contract
+  `AgentStream` parses and the two behaviours an adapter most easily breaks (the read cap and
+  the cost ledger), the result-envelope requirement with its fail-closed rule, and the five
+  test files a PR would have to touch. Linked from the README.
+
+- **`tldrx questions cards [<run>]` — a parked question, as something to decide (#59).** Measured
+  on run `260830-ordering-inventory`, 2026-09-01: the host parked four product questions with
+  notes and reported them in its tl;dr. The owner's live words were *"cuales preguntas? no las
+  veo? por que no me guió por las preguntas Claude?"* — they were on disk, they were in
+  `tldrx questions <run>`, and the count was in the summary. None of that PRESENTED them. Counted
+  is not asked, and the gap was never the data; it was the arc from parked to answerable.
+  - **Three slots, each a refusal.** Two lines of context (which run and file the question is
+    parked in, who asked it, when, in what area) so a card pasted into chat stands up away from
+    the terminal. The question's OWN `Why asked:` note verbatim, `[src: …]` included — the slot
+    for what the binding docs already decide, quoted rather than summarised; a note that cites
+    nothing is FLAGGED as somebody's recollection, and a question parked with no note says so.
+    The file's lettered options verbatim, or a loud `NEEDS OPTIONS` marker when it carries none,
+    because manufacturing A/B/C would be answering the question in the act of asking it.
+  - **A reader, and only a reader** (owner decision 2026-09-01: no interactive loop in v1). It
+    opens no run, spawns nothing, records no fact, and a test asserts `questions.md` is
+    byte-identical across a render. Answers still flow through `tldrx answer`, and every card
+    prints the exact line to type. It does not extend the §2.7 grammar: every field on a card is
+    one the existing parser already produces.
+  - **No open question is a sentence and an exit 0** — and two silences are told apart, because
+    "this run never parked anything" and "everything here is answered" send a reader to different
+    places.
+
+- **`tldrx watch arm --run <id>` — the merge detector that fires the post-merge checklist (#69).**
+  `watch check` (#65) answers *"what do I check now?"*. Nothing answered *"the PR just merged, go
+  and check it"* — the half that happens without a human remembering, which matters because the
+  failure #65 was filed about IS a memory failure: the owner's own CD gap, a merged branch read as
+  a deployed one, cost 19 destroyed records. v1 would not have caught it either, because nobody
+  ran it.
+  - **A bounded FOREGROUND poller, not a daemon** (owner decision 2026-09-01: no GitHub Actions in
+    v1). It reads the branch Build cut from `run.yml` — through the same `pickBranch`/`findRepos`
+    `tldrx ship` uses, now exported, so the two verbs cannot disagree about which branch a run
+    shipped — asks `gh pr view <branch> --json state,mergedAt` per repo, and prints the same
+    checklist the moment every PR for that branch has merged. **Three independent bounds**: a hard
+    deadline (`--timeout`, default 3600s, max 86400), a floor on the interval (`--interval`,
+    default 60s, under 10 REFUSED rather than quietly raised), and a poll cap that holds even if
+    the clock does not move.
+  - **Every refusal is a sentence, and they are different sentences.** No epic branch ⇒ `ship`
+    cannot have run. No PR for the branch ⇒ either `ship` has not run or the branch was never
+    pushed, with both commands printed. A PR `CLOSED` without merging ⇒ stop now, because waiting
+    for it could only ever time out. A window that expires ⇒ exit `4` and the command that
+    re-arms it.
+  - **It never pushes, opens or merges anything**, and `--execute` is deliberately not offered: an
+    hour-old poller must not start running the workspace's build commands the instant a merge
+    lands. Re-running a recorded command stays an explicit, typed decision.
+  - **No test in this suite runs the real `gh` or touches a network.** The unit cases drive a
+    recording fake transport; the one end-to-end case puts a stub `gh` first on `PATH`. The clock
+    and the sleep are injected, so a test covering a one-hour timeout finishes in milliseconds.
+
+- **`tldrx plan schema` — the story/epic/waves contract, printed for a human (#71).** #48 deleted
+  `templates/story.md` and `templates/epic.md`, rightly: nothing read them, so nothing kept them
+  honest. But they were answering a real question — what shape does a story file take? — and after
+  #48 the only reader of the generated contract was `checkContracts.ts`, which splices it into the
+  Plan stage's prompt. The shape existed for the agent and not for the person writing a story by
+  hand, reviewing one an agent wrote, or debugging a `plan` check refusal.
+  - **The same bytes the agent gets.** `plan schema` renders `renderPlanSchemaContract()` verbatim,
+    generated from the validators the check runs, so it cannot become the second source of truth
+    the templates were. `--story`, `--epic` and `--waves` print one example on its own; at most one
+    of them, because two would make the answer ambiguous. A test copies the printed story example
+    straight into `validateStoryFile` and asserts it comes back clean.
+  - **The one verb in `plan` that resolves nothing.** No workspace, no run, no disk, no spend — the
+    question comes before any of that exists, and is often asked from outside a workspace entirely.
+
+
+- **`tldrx drive [--attended|--unattended]` — the host/driver mandate, shipped (#63).** Every elite
+  run of 2026-08-31/09-01 was driven by a session carrying a hand-written playbook, and that playbook
+  was the framework's real quality floor. It lived in the owner's chat pastes, so a third party
+  inherited the CLI and rediscovered the discipline — or did not. It is now plain text the package
+  prints, versioned with the package, for a human to paste into the driving session or read before
+  they start.
+  - **What it carries**: the three-role protocol (developer sub-agent → a FRESH adversarial reviewer,
+    never the author → the host verifying BOTH in the code, not in their reports); evidence
+    discipline (measured / inferred / assumed labelled in the same sentence as the claim, exit codes
+    never read through a pipe, verification from the source, remote shas via `git ls-remote`);
+    parking product questions with what the docs already decide, because an open question is never a
+    licence to ship an unguarded write; review calibration by stakes, so a security-bearing story
+    gets the strongest reviewer; and budget honesty — declare a turn once, state a floor when the
+    records are incomplete.
+  - **Two modes, one spine.** The disciplines are identical — they are about how a claim is made,
+    not about who is watching. What differs is the GATE (`--unattended` signs over a validated
+    evidence note with `tldrx approve --as-agent`; `--attended` never signs and hands the decision
+    over) and DRIVING (who spawns). Each mode is bounded at 120 lines, asserted by a test: a mandate
+    long enough to skim is one nobody follows.
+  - **A mode is required and never guessed** (exit `1`), the same refusal `tldrx run attend` makes.
+    Handing an attended session the unattended text tells it to sign gates that were never its to
+    sign. It needs no workspace, opens no run, spawns nothing and writes nothing.
+
+- **`tldrx retro --all` — cross-run mining of what keeps catching you (#64).** Each run left gold
+  nobody aggregated: a reviewer's verdict and findings in `04-build/log/<story>.md`, a fix list with
+  a disposition per finding, `retro.md`'s `## Build feedback`, and the reason a person typed when
+  they reopened a story. Across the first six real runs the same finding CLASSES kept coming back,
+  and nothing read more than one run at a time.
+  - **A trends table**: finding class × count × how many runs it appeared in × one example with the
+    `[src: …]` that lets a reader go and check it.
+  - **Seven classes, in a documented precedence**: `test-cannot-fail`, `missing-negative-control`,
+    `unreachable-structure`, `stale-comment`, `authorization-not-widened`, `schema-drift`, `other`.
+    Classification is ordered keyword rules over the finding text — no model, no scoring, no
+    threshold — so the same tree always produces the same table and a rule that misfires can be
+    pointed at. The rules are tested against fixtures in the shape the real artefacts carry.
+  - **Zero new state.** It writes nothing anywhere: no `retro.md`, no `practices.md`, no cache.
+    A test asserts the workspace is byte-identical across the call.
+  - **Absence is never an error.** A run with no Build phase, no retro, no events log or an
+    unreadable one contributes what it has and is still counted; an empty workspace is an empty
+    answer at exit `0`. A repeat of one finding WITHIN a run is collapsed (`retro.md` quotes the fix
+    list verbatim); the same finding in two runs is two occurrences, which is the point. A `refuted`
+    fix-list finding is read and dropped — ranking a class by disproven findings would make the
+    table a report on the reviewer.
+  - `--all` is refused (exit `1`) alongside a `<run-id>` or `--apply`, before a file is opened: each
+    asks for the opposite of what `--all` does.
+- **`tldrx watch check` is the post-merge checklist (#65, owner decision 2026-09-01: manual command
+  first, `gh` detector later).** `tldrx ship` opens the PR carrying the handoff and the watcher card
+  lists the signals that would prove the feature works — and nothing joined the two, so
+  "merged ≠ deployed ≠ verified" stayed a thing a person had to remember. Remembered wrongly once, it
+  destroyed 19 records. `watch check` now reads the run's cards under `05-watch/watchers/` and prints
+  each card's `## Signal` items as a numbered checklist with `## Where`, the baseline, what broken
+  looks like, and the Query block, so there is one screen to work through after a merge.
+  - **A feature id is now OPTIONAL.** `tldrx watch check` checks every card in the run — the shape a
+    CI job wants; `tldrx watch check <feature>` scopes it to one and still prints the citation
+    verdict `check` always printed. The citation re-check is unchanged and still exits 1 on a dead
+    `[src: …]`.
+  - **Owners are DERIVED, never invented.** The watcher schema has no `owner:` key and adding one
+    would have been a key nothing writes (the #48 defect again). The owner of a signal here is the
+    repo its own citation names — `[src: api:src/Leaderboard.cs:64]` says `api` — and an item that
+    cites no repo says so rather than borrowing the card's.
+  - **Runnable means the card AND the workspace say so.** Only a `$ <cmd> → exit <n>` source whose
+    command `.tldrx/workspace.yml` declares, in a repo the item or the card names unambiguously, is
+    offered. `--execute` (off by default) re-runs exactly those, through `runDeclaredCommand` — the
+    same allowlist, argv-never-a-shell and timeout the stage `cmd` check uses, extracted rather than
+    copied — and reports the exit each gets NOW against the exit the card recorded. That is the one
+    thing in the framework that catches a card whose `$ … → exit 0` has quietly become an exit 1:
+    `resolveSrc` checks a `cmd` source's MEMBERSHIP and takes the exit code on the agent's word.
+  - **A `## Query` block is never runnable.** It is KQL or SQL for the console named under
+    `## Where`, which tldrx has neither credentials nor a client for, so it is reproduced with its
+    fence language and marked print-only.
+  - **Three refusals, told apart because they need different actions.** No `05-watch/` at all ("its
+    Watch stage never ran"), a Watch stage that wrote no card ("no story reached `done`"), and a
+    `draft` card — which is not a failure but an answer: the card's own `absent:` sources are quoted
+    back as what to instrument. The two empty cases exit 3, never 0; a green meaning "I read no
+    cards" is the failure this command exists to stop.
+
+- **Evals v1 — five golden-transcript evals, one per stage (#26, owner decision 2026-09-01: small
+  v1 now).** The suite proved the harness; nothing proved a STAGE. `test/evals/` now runs each of
+  What, How, Plan, Build and Watch through the real facilitator against the scripted stand-in agent
+  and asserts that stage's output CONTRACT. About 5 seconds on top of `bun test`.
+  - **One stage at a time, not a chain.** Stages are sequential — Plan cannot run until How's gate
+    is signed — so playing four stages to reach the fifth would make every eval depend on the ones
+    before it, and a regression in What would turn all five red without any of them saying why.
+    Each eval instead opens its own run on a workflow preset holding exactly one stage, which needs
+    no new code: `workflowPath` already prefers `.tldrx/workflows/<scope>.yml` over the shipped
+    presets, and `normalisePhase` already takes the phase folder from the stage's own `phase:`
+    rather than from its position. Whatever the stage genuinely reads — a plan for Build, a done
+    story for Watch — is seeded onto the disk by the scenario.
+  - **Built on `learn`'s machinery, not beside it.** The stand-in `claude`
+    (`src/core/learn/agentScript.ts`, `learnAgent.ts`) and the toy-repo sandbox are reused as they
+    are; the real CLI is unreachable by the same two doors `learn` closes. No production code
+    changed for this.
+  - **Contract, not snapshot.** Nothing compares bytes. The evals assert what the FRAMEWORK
+    computes: the declared outputs read back off `gate.requested`, the `checks:` outcomes and their
+    computed detail (`checkPlan`'s branch-model line), the artifacts re-parsed with the framework's
+    OWN validators (`validateQuestions`, `parseWatcherCard`), and the side effects a stage exists
+    for — a branch cut, the story's `dod` block re-run and its exit code recorded, a `--no-ff` merge
+    into the epic branch with `main` untouched, and a watcher card's status COMPUTED as `verified`
+    when the scenario's front matter claimed `draft`.
+  - **A check that is `skipped` for any reason other than being write-time-only now fails an eval.**
+    `runCheck` falls through to `unknown check id '<x>'`, which is what a renamed or deleted check
+    looks like from outside — and it would otherwise read as a pass.
+  - **`EVALS` is load-bearing, not bookkeeping.** A coverage test asserts it names every stage
+    under `stages/`, so a sixth stage shipping without an eval turns the suite red and says which
+    one is missing.
+  - **Every eval was watched failing before it was trusted.** One sabotage each, listed with its
+    symptom in `test/evals/README.md`; the Plan one is the shape to aim for — nothing crashed, every
+    count still matched, and dropping a cross-epic `depends_on` was still caught because the derived
+    branch model changed. The same file says how to add a sixth eval, and what v1 deliberately does
+    not cover: failure paths, the agent gate, attended runs, and prompt QUALITY, which needs a real
+    model and a judge.
+
+- **`tldrx story reopen <id> --for-fix --note "<defect>"` — a sanctioned fix round on a `done` story
+  (#58, owner decision 2026-09-01).** Measured on `260829-scoring-leaderboard`: S11's adversarial
+  review found a real defect (linkEmail succeeds, setDisplayName fails, the account is permanently
+  linked and the score never claimable). It was accepted, small and well understood — and the story
+  was already `done`, so the only choices were rejecting the whole Build stage, which destroys
+  fourteen good stories' closure, or fixing it outside the story machinery, which leaves an
+  epic-level commit with no story provenance. `done` → fix round was the missing arc.
+  - **No attempt is consumed.** The mechanism is the one that was already there: `story.reopened` is a
+    reset boundary the review ledger reads, so the approve that closed the story stops counting and
+    the fix runs as attempt 1 of 2 — with both turns available to it.
+  - **The same DoD and the same reviewer.** The story goes back to `todo` and the Build pipeline picks
+    it up unchanged; a fix a reviewer refuses twice blocks, exactly as the original would have.
+    Nothing is waved through because a human asked for it.
+  - **It cannot relitigate scope.** The `--note` names a defect, and `status:` is the ONLY line the
+    verb moves on the story file — the acceptance criteria the reviewer will judge against are the
+    ones that were already there.
+  - **One open fix round per story.** It opens on a `story.reopened` carrying the new `reason: fix`
+    and closes when the story is `done` again; a second `--for-fix` while one is open is refused,
+    naming who opened it and with which defect. The bound is read from `events.jsonl`
+    (`ReviewLedger.fixRound`), so it holds across processes.
+  - Refused when the story is **not** `done` (an unfinished story is the plain verb's job), when
+    `--note` is missing, and on that second round. `reason` is written on every `story.reopened` now —
+    `fix` or `attempts` — and an event without it predates the key and is an `attempts` reopen, the
+    only kind that existed.
+  - The plain verb's `done` refusal now points at `--for-fix` beside `reject --stage`.
+
+### Changed
+
+- **tldrx is no longer sold as "lightweight" (owner decision, 2026-09-02).** The word survived from
+  the concept doc, where it was true of a proposal. What it now describes is five stages, a
+  machine-enforced `[src: …]` grammar that refuses a write it cannot resolve, an adversarial
+  reviewer that is never the author, a dual-economy budget ledger, and a written trail from `run
+  new` to `retro`, held up by a suite that measured 3093 tests across 108 files on 2026-09-02. One line replaces it everywhere, so
+  the npm page, the binary's own `--help`, the plugin manifest, the README and both landing pages
+  say the same thing: **an evidence-first, file-based AI development framework: five stages, a gate
+  on every one, and every claim cited or refused.**
+  - Changed in `package.json`, `src/cli/commands/help.ts` (the first line of `tldrx --help`),
+    `plugin/.claude-plugin/plugin.json`, `plugin/skills/tldrx/SKILL.md`, `docs/concept.md`,
+    `README.md`, and the docs site's `description`, hero and Spanish mirror.
+  - **The status tag moved with it, because it was already stale.** 0.4.0 shipped as `beta`
+    (`CHANGELOG.md`, `README.md` release table) while the badge, the README preamble, the binary,
+    the plugin manifest, the site's "Where this is" section, both site footers and the generated
+    release-notes preamble all still said `alpha`. They say `beta` now. The site's beta bar is
+    stated as CLEARED rather than as being worked through, which is what the release table has
+    claimed since 0.4.0.
+  - `docs/RELEASING.md` gains the thing that made this drift invisible: the `status-…` badge is a
+    hardcoded shields.io URL and does **not** update on its own, unlike the npm and CI badges
+    beside it. The README's own release paragraph now shows `--tag beta`, because `release.sh`
+    writes `alpha` when `--tag` is omitted.
+
+- **A watcher card may name a HUMAN owner — optional, per item, never invented (#70).**
+  `watch check` derived an owner from each item's own citation: `[src: api:src/Leaderboard.cs:64]`
+  → `api`. That was right for v1 and stays the fallback, but it answers a different question from
+  the one #70 asked — which repo EMITS a signal is not who gets paged when it stops.
+  - **Additive, in both halves.** An optional front-matter `owner:` (validated only when present,
+    so every card already on disk still validates) and an optional `(owner: <name>)` annotation on
+    an individual item, placed BEFORE its `[src: …]` token because §2.8 makes that token the last
+    thing on the line. Resolution is item → card → repo-derived, and the printed line says WHICH
+    it is showing: `owner: alice (declared on the item)` versus the pre-#70 `owner: api`, left
+    byte-identical for every card that declares nothing.
+  - **Filled from the ledger that already names owners, or not at all.** `tldrx init` parks
+    "Who owns `<repo>`?" as an `ownership` question and the answer lands in
+    `.tldrx/memory/facts.yml`. The Watch prompt was inlining `observability` and `deploy` facts
+    only, so a sub-agent asked for an owner had no honest source and would have invented one;
+    `ownership` is now inlined, and the brief says the name may come from nowhere else.
+  - **A lost name is an error, not an absence.** `(owner: )` is a card that TRIED to name somebody
+    and lost it, so it is a shape issue on the card rather than a silent fall-through to the repo
+    — which is the exact substitution the issue is about.
+
+- **`docs/guide/08-cli-reference.md` documents `note` and `ship` (#72).** Every command in `COMMANDS`
+  had a `## tldrx <cmd>` heading except three; #55 wrote `plan`'s, and these are the other two —
+  `ship` being the command that opens the PR at the end of a run, and `note` how an operator records
+  something against a stage. `DOCUMENTED_SUBCOMMANDS` in `test/cli.test.ts` grew to cover them, which
+  is what keeps each of these gaps a red test rather than a note. Only `hook` is left out, and
+  deliberately: its seven scripts are documented as the one `<script>` slot `USAGE_SPELLINGS` already
+  records as a spelling.
+
+- **`tldrx watch`'s one-line summary says checklist.** It read "List and re-check the watcher cards a
+  run produced", which described half of what `watch check` is for and disagreed with the `--help`
+  text #65 updated.
+
+- **`tldrx ship` opens one PR PER REPO when the branch is in more than one (#66, owner decision
+  2026-09-01).** Since #57 a chained multi-repo run cuts ONE integration branch, `epic/<run-id>`,
+  with the same name in every repo — so `ship` found it in several, every time, by construction, and
+  refused with `pass one: --repo <name>`. The last step of every such run was typing the same command
+  once per repo and remembering which ones had already gone through.
+  - Same handoff as the body of every PR, the repo name in the title, and every URL listed at the
+    end. Each PR opens against **that repo's own** `default_branch`.
+  - **One repo is byte-identical**, down to the four lines it prints and the number of processes it
+    spawns: the `gh pr list` probe below exists for the multi-repo case and never runs when there is
+    only one. The single-repo lines are asserted as exact strings, not substrings — "we did not
+    change the common case" is not a claim a `toContain` can make.
+  - **A partial failure names both sides.** PR 2 of 3 failing still opens PR 3, and the report lists
+    the repos that succeeded with their URLs and the repos that failed with the reason (exit `2`).
+    Aborting on the first failure would leave a half-shipped run and no statement of which half.
+  - **Re-running is safe.** Before creating, each repo is asked whether an open PR for the branch
+    already exists (`gh pr list --head`); one that has is skipped and listed. So the fix for a
+    partial failure is `tldrx ship` again and nothing else. A `gh` that fails or answers with
+    non-JSON is treated as "there is none", so a transient error can never silently turn a real ship
+    into a skip.
+  - `--repo` still narrows to exactly one, and `ship` still never pushes: an unpushed branch in one
+    repo is that repo's failure line, naming its `git push`, while the other repos' PRs still open.
+
+### Fixed
+
+- **A training pass with nothing to read is no longer paid for (#101).** `runTraining` had
+  exactly one "nothing to work on, refuse before the money" check and it guarded one half of one
+  mode: `nothingToMineRefusal` fires only for a ROLE expert with zero minable runs. The CODE pass
+  was pushed unconditionally and nothing ever looked at `selection.inlined.length`, so an expert
+  whose `## Domain` matched no file on disk spawned a sub-agent, was shown no code at all, and
+  wrote a knowledge file about nothing at full price. Measured on the training fixture before the
+  fix, with a domain of `src/does-not-exist/`: **`{ code: 0, costUsd: 0.37, wroteKnowledge: true }`**
+  — a successful run, priced, from zero input. Live cost, from #94's thread: two near-empty
+  trainings at $0.82 each whose "code sweeps found nothing in-domain".
+  - **The choice, named: SKIP a dead pass, REFUSE only when none survives.** That is
+    `nothingToMineRefusal`'s own idiom — refuse when there is nothing to train from — without
+    denying a `--mode full` run that still has one real pass. The skipped pass's reason goes to
+    stderr before the money, exactly like the #96 pre-start line, and rides back in
+    `TrainOutcome.warnings`; `--prepare` prints it too, and its `N sub-agent(s)` line was already
+    derived from the surviving passes so it stays honest by construction.
+  - **`roleTraining.ts:79`'s uncovered arm is closed.** Its guard reads
+    `if (!isRole || minedFiles > 0) return null`, so a NON-role `--mode full` run against zero
+    minable runs was never refused: it spawned a second sub-agent to write
+    `- none [src: absent:tldrx-work]` — no evidence, no level, full price. That case is now a
+    skip. The role case keeps its own, better-worded refusal untouched, because its runs pass is
+    the only pass it has.
+  - **It is NOT the #96/#98 preflight and does not touch it.** Underfunded and empty are different
+    refusals with different remedies, so they are different checks that never consult each other.
+    In particular a skipped pass does not re-divide the ceiling: the share `trainPreflight` priced
+    and printed is what the surviving sub-agent gets, and the skipped one is simply not spent.
+  - **Exit `1`, deliberately.** `2` is this codebase's MONEY refusal (the `MIN_TRAIN_USD` floor and
+    the #96 preflight). Every "you asked for something with nothing behind it" refusal in
+    `expert train` is `1` — `missingAreaRefusal`, `lightModeRefusal`, and `nothingToMineRefusal`,
+    this check's literal sibling, which `test/training.test.ts` pins as "refused (exit 1)".
+  - **One doc correction found while checking that.** `docs/guide/08-cli-reference.md` claimed
+    `--mode light` on a role expert exits `2`. It exits `1`: `lightModeRefusal` returns
+    `EXIT_USAGE`, and `src/cli/exitCodes.ts` defines `EXIT_USAGE = 1`.
+
+- **`expert create` now yields an expert that can actually be trained, and states the `## Domain`
+  grammar it will be read under (#94).** Live 2026-09-02: `tldrx expert create discoverer` printed
+  *"no areas — every level starts at 0"*, and `tldrx expert train discoverer --area discoverer`
+  answered *"has no area (areas: none)"*. Neither sentence said HOW to add one; no flag or
+  subcommand existed; the sanctioned path — a block in `competencies.yml` — was discoverable only
+  by reading the source. Then the same gap cost money twice more: the hand-written `## Domain`
+  bullets came out WORKSPACE-relative (`Scavtopia.Workflows/src/…`) because nothing documents the
+  grammar, so **a $2.10 full training earned zero evidence — all 13 of its code citations read
+  `outside domain`** — plus two near-empty $0.82 runs whose code sweeps found nothing in domain.
+  - **`--area <id>` seeds the first competency area, and `--title <text>` names it.** The title is
+    not decoration: light mode greps the words of the area title to choose which files the expert
+    is shown (`training/selectFiles.ts`), so `create` prints the default title it chose and names
+    `--title`. An area id obeys the same slug rule the expert name does — it is a filename
+    (`knowledge/<area>.md`) and half of a copy-pasteable `--area` argument.
+  - **Both refusals name the FILE and print the block.** `expert train` on a missing area now
+    answers with `.tldrx/experts/<name>/competencies.yml` and the five keys to paste (`id`,
+    `title`, `level: 0`, `train_prompt`, `evidence: []`), from one helper both the CLI and
+    `runTraining` use, so the two paths cannot drift. `create` with no area says the same thing.
+  - **`create` writes the front-matter `repos:`**, read off `.tldrx/workspace.yml` (`repos: []`
+    when there is none, rather than a guess). `expert.md` had no such key at all, so a
+    hand-created expert declared nothing about which repos its `## Domain` bullets are relative
+    to — and that is exactly the thing the bullets are relative to.
+  - **The created `expert.md` states the grammar in its own `## Domain` section**: bullets are
+    repo-RELATIVE with no repo prefix, a whole-repo claim is `` - repo `api` ``, citations arrive
+    as `repo:path:line` so `api:src/Checkout/Cart.cs:12` is matched by `` `src/Checkout/` `` and
+    **not** by `` `api/src/Checkout/` ``. It is written as prose, deliberately: a worked example
+    written as a BULLET would itself be parsed as a declared domain path, and an expert that
+    silently claims `api/src/Checkout` puts every real citation outside its own domain — the bug,
+    not the fix. Pinned by a test that the created file still parses to zero domain paths.
+  - **The `outside domain` warning names the grammar FIRST** when the cited path would match with
+    one leading segment dropped from either side — the bullet carrying a repo prefix it should
+    not, or the citation being workspace-relative where a `repo:path:line` was wanted. It says the
+    measured fact (`X` without its first segment DOES contain this path), not a diagnosis.
+  - **And naming another expert no longer reads as exclusivity.** *"train that expert on it
+    instead"* was measured false: overlap between domains is legal, and the hint fires whenever
+    this expert's own paths miss. It now reads *"also declares a domain that contains it, so it
+    may be the better home for this one (overlap is legal: this fired because none of `<expert>`'s
+    own paths match)"*.
+  - **`init/loadWorkspaceFile.ts` reads `.tldrx/workspace.yml`'s name from `core/paths.ts`**
+    (`PROJECT_WORKSPACE_FILE`, the constant #92 added) instead of from `init/runInit.ts`;
+    `runInit` keeps exporting `WORKSPACE_FILE` under that name, aliased to the same string, so no
+    import site changed. `loadWorkspaceFile` is the small leaf several commands read the repo list
+    through, and taking that one constant off `runInit` dragged the whole of `init` — detection,
+    map providers, the MCP probe — into every module that read it: `dist/hooks/session-start.js`
+    went 37,937 → 54,467 bytes and blew the 50 KB entry-point cap `test/build.test.ts` enforces.
+    It is back to 37,937.
+  - Documented in `docs/spec.md` §2.3 and §2.8, `docs/guide/04-experts.md`,
+    `docs/guide/08-cli-reference.md`, `docs/guide/09-troubleshooting.md` and both docs-site expert
+    guides — the grammar had never been written down anywhere outside `expertDomain.ts`.
+
+- **`release.sh` ran the gate AFTER pushing the release commit to `main` (#100).** The order was
+  commit → push `main` → `release-check.sh` → tag → push tag, so every red item the gate has —
+  tests, typecheck, build, the `Bun.*` seam grep, "tag already exists", "already on npm", "tree
+  not clean" — landed a `release: X.Y.Z` commit on `origin/main` carrying a **dated** CHANGELOG
+  heading and a **dated** README row, with no tag behind it. That is precisely the half-released
+  state checklist item 4 exists to prevent, and it costs a revert commit on `main` or a
+  hand-repaired CHANGELOG to undo. The order is now commit → **gate** → push `main` → tag → push
+  tag: a red gate leaves `origin/main`, the tags and npm untouched, the whole of the damage is one
+  local commit, and the script says so and prints the one command that drops it.
+  - **The gate did not have to be weakened to move it.** Of its items only "in sync with
+    `origin/main`" assumes the push already happened; run against an unpushed release commit it
+    would be permanently red, which is the trap in the naive reorder. `release-check.sh` takes a
+    new `--pre-push` that restates that one item as **"`origin/main` is HEAD's parent"** — the same
+    assertion (nobody moved `main` under you, nothing but the release commit is unpushed) for a
+    tree that has not pushed yet, and one notch stronger, since it also refuses a second unpushed
+    commit. Items 1–3 already needed the edits, which exist by then; "working tree clean", "on
+    main", "tag does not exist", "not on npm", typecheck, tests, build and the seam grep are
+    untouched and all still run. **No flag, no default, no CI path changed**: bare
+    `release-check.sh` (the `release-gate-hook.sh` PreToolUse deny) and `--ci` (publish.yml)
+    behave exactly as before, and there is a test for each.
+  - **`sed -i ''` was macOS-only**, which is why this path had never had a test: measured against
+    GNU sed 4.9 in `debian:stable-slim`, `sed -i '' -E …` exits **2** with `sed: can't read
+    s/^## …/: No such file or directory`, so `release.sh` could not run on the ubuntu CI runner at
+    all. Both substitutions now go through a temp file outside the tree and are `cat` back (inode,
+    mode and a clean working tree preserved) — byte-identical output on BSD sed, and it runs on
+    GNU sed, which is what lets `test/release-gate-order.test.ts` hold the ordering in CI. The new
+    file runs the real two scripts against a real bare "origin" with only `bun` and `npm` stubbed,
+    and reads the ordering off the shas the gate actually ran against: pre-fix, HEAD and
+    `origin/main` were EQUAL at gate time; now they differ by the release commit.
+
+- **The prose was validated against the binary, top to bottom, and a lot of it was false.**
+  The last docs QA (3ee3723) was a release ago; `watch check`/`watch arm`, `update` and
+  its notice, the `drive` preflight, `plan schema`, `questions cards`, `story reopen --for-fix`,
+  the host-token economy and the dashboard's new sources have all landed since. Every claim in
+  `README.md`, all thirteen English pages, all thirteen Spanish pages and `CONTRIBUTING.md` was
+  cold-read for lies first, then checked against `tldrx <cmd> --help` and the source. What was
+  found was not drift at the edges — most of it was load-bearing.
+  - **A flag that does not exist.** The README's cost section explained what `--max-budget-usd`
+    does. That is the flag tldrx passes DOWN to `claude`, not one a user may type: `tldrx next
+    --max-budget-usd 5` answers `unknown flag`. The user-facing flag is `--max-usd`.
+  - **"Every run-targeting command exits `2`" was wrong twice**, in the README, the CLI reference
+    and the FAQ. `tldrx cost` refuses at exit **1** (`cost.ts` returns `EXIT_USAGE`, and its own
+    `--help` never lists a 2), and `tldrx run status` does not refuse at all — it lists every open
+    run and exits **0**, which is the whole point of it, since it is the screen you read to find
+    the id the others are asking for. All three pages also named five commands as taking a
+    positional `<run>`; the real split is that most take either form, `replay` and `retro` take
+    the positional only (`--run` there is an unknown flag, measured), and seven take `--run` only.
+  - **A documented command that silently creates a directory where you asked for a file.**
+    `tldrx dashboard --static --out ./somewhere/page.html` was in the reference. `--out` is a
+    directory and the filename is not negotiable (`writeStatic.ts` joins `index.html` and
+    `mkdirSync`s the path), so that line makes a directory literally named `page.html`.
+  - **A required tool the Quickstart did not name.** `env.yml` declares four `required: true`
+    tools; the install page named Node, `git` and `claude` and never mentioned **Bun ≥ 1.3**, so a
+    reader who followed it exactly got `tldrx doctor` exit 1 and no explanation.
+  - **The `tldrx learn` chapter list was stale AND swapped** on both language Quickstarts: 6 and 7
+    are `the agent gate` and `attended`, and the pages had two other titles in the other order —
+    so `--chapter 7` sent a reader looking for the agent gate to the wrong chapter.
+  - **Counts that had grown.** The `.gitignore` block `init` writes excludes **eight** paths, not
+    the five the README listed. The `[src: …]` grammar has **eight** kinds, not the seven the
+    evidence table showed — `aidlc:` was missing while the page's own next paragraph said
+    "eight". The What, How and Plan stages each declare `questions.md` as an output and Plan also
+    declares `budget.yml`; the stages table showed none of them, which matters because the auto
+    gate is measured on exactly that file. `tldrx init` seeds **three** kinds of expert — role,
+    stack and **domain** — and both expert sections said two.
+  - **Twelve `workflows/*.yml` said the auto gate has "the five spec §5 conditions"** and listed
+    five. `evaluateAutoGate` has had **seven** since the stories and boundary conditions landed.
+  - **The `tldrx drive` mandate's preflight (#84) was absent from every page that enumerates what
+    the mandate carries**, in both languages, and the English driving guide still taught the
+    three-command recipe the preflight replaced.
+  - **`CONTRIBUTING.md` overstated the ref guard and understated the lock.** The guard refuses a
+    ref move while ANOTHER invocation holds the lock — the holding run's own git children carry
+    `MW_LOCK_TOKEN` and pass, which is what lets the wave commit and push through its own guard;
+    as written it would have refused the merge it exists to protect. `merge-wave.sh`'s seven exit
+    codes were documented nowhere, in a file whose own rule is to read exit codes. The lock's wait
+    is bounded (`MW_LOCK_WAIT_S`, default 3600 s, then exit 6) and `CLAUDE.md` implied it was not.
+  - **The Bun/Node seam rule was stated more broadly than it is enforced**, in `CONTRIBUTING.md`,
+    `CLAUDE.md` and `docs/RELEASING.md`: the grep scans `src` only, and `scripts/build.ts` calls
+    `Bun.build` for a living. A contributor reading it literally files a bug against the build.
+  - **`publish.yml` does not "run the same check".** `release-check.sh --ci` skips items 4 and 5
+    entirely; the seam check runs in no CI workflow at all. Said plainly now, in both the README
+    and `docs/RELEASING.md`.
+  - **Four translated strings that had to stay English** in the Spanish mirror: two `--note`
+    samples, the `## Qn · Title` heading the §2.7 parser matches on, and `$0.00 of $25.00`, which
+    is what the dashboard actually renders.
+  - Fixed on the way: the `0.5.0` CHANGELOG section had grown **two `### Fixed` groups** and had
+    two additions (Evals v1, `story reopen --for-fix`) filed under `### Removed` — an artefact of
+    unioning sibling branches. One group per kind now, in the order `0.3.0` used, with every entry
+    byte-identical and only the duplicate heading gone.
+  - Two defects found while checking and **filed rather than fixed**, both outside this pass:
+    [#99](https://github.com/ederwii/tldr-experts/issues/99) `test/merge-wave.test.ts` asserts over
+    a shared `TMPDIR`, so a sibling process fails it (measured: red under load, green alone, and
+    green for the full suite with no sibling), and
+    [#100](https://github.com/ederwii/tldr-experts/issues/100) `release.sh` pushes the release
+    commit to `main` before it runs `release-check.sh`.
+
+- **`expert train` says what it is about to spend, on which model, BEFORE it spends it (#96).**
+  Live 2026-09-02: `tldrx expert train discoverer --area discoverer --mode full` inherited the
+  claude CLI's last-used model — `fable-5`, a premium tier — and ran against the default ceiling.
+  Full mode splits that ceiling between its two sub-agents, so the code pass was handed $1.00. It
+  died with `Reached maximum budget ($1)` at 54 s, **$1.31 spent and recorded, nothing written to
+  `competencies.yml`.** Three defaults compounded and not one of them was said out loud.
+  - **A pre-start line names the model, its tier, and where the name came from** — `--model`,
+    `$ANTHROPIC_MODEL`, or a `model:` key in `.claude/settings.local.json` /
+    `.claude/settings.json` / `~/.claude/settings.json`: `model claude-fable-5[1m] (premium,
+    inherited from your claude CLI via ~/.claude/settings.json) — pass --model to override ·
+    --mode full · $3.00 across 2 sub-agent(s), $1.50 each`. When nothing on the box says which
+    model the CLI will pick, the line says THAT rather than inventing a tier.
+  - **A refusal — exit `2`, nothing spawned, nothing spent — when the share cannot fit and the
+    ceiling is the DEFAULT one.** The test is arithmetic, not a category: does the money one
+    sub-agent gets reach what one pass on that tier costs? Measured full trainings run
+    **$1.21–$1.60 end to end on a mid model** (two `training.jsonl` lines on `aparece-platform`,
+    `docs/audits/2026-08-29/experts-knowledge.md` §E, plus the top of the band in #96), so one
+    pass is ~$0.70 mid and ~$1.76 premium (opus lists at 2.5x sonnet). Full mode's $1.50 a pass
+    does not reach it. The refusal names both remedies — `--model sonnet`, or `--max-usd <n>`.
+  - **An explicit `--max-usd` is never refused.** The operator looked at the number; that is the
+    whole decision this check exists to ask for. It warns in one line and proceeds.
+  - **The full-mode default `--max-usd` is now $3.00**, up from the $2.00 light mode still uses,
+    because full mode pays for two sub-agent passes and the one repair round a rejected knowledge
+    file earns comes out of the same share. $1.50 a pass is ~2x the measured per-pass midpoint.
+    Deliberately not scaled higher for a premium model: `--max-budget-usd` is a stop after the
+    turn, not a cap (spec §2.6.1 — a $1.50 ceiling has realised $5.15), so a bigger default cannot
+    make a premium turn affordable, only more expensive to lose. `budget-gate` prices `--mode
+    full` at $3.00 to match.
+  - Salvaging the partial evidence a budget death throws away is a design call and was NOT built;
+    the question is asked on #96.
+
+- **`expert train --prepare` carries that check into the bundle (#98).** The check above was wired
+  to the spawning path only, and `--prepare` spawns nothing — but it writes the ceiling into
+  `pending.json` (`max_budget_usd: 1.50`) and into the prompt text (`Ceiling for this sub-agent:
+  $1.50`), and a host session then spends against it. Measured on `b5d59c5`: a full `--prepare` with
+  a premium model inherited from `~/.claude/settings.json` printed `$1.50 ceiling each`, returned
+  `preflight: null`, and wrote two bundles saying `model: null` — the same trap, one command later,
+  on the host's money.
+  - The model line and any warning now go to **stdout with the prepared block** and into an optional
+    **`preflight`** key on each bundle's `pending.json`. Absent when there is no warning, so an
+    unremarkable bundle stays byte-identical to the one this command has always written.
+  - **What it may CLAIM differs by half.** Headless, tldrx spawns `claude` with no `--model`, so the
+    CLI's default is a prediction about a process this code starts: the refusal stands. On
+    `--prepare` an **explicit `--model`** is an instruction written into the bundle and is refused
+    the same way (exit `2`, no bundle written); an **inherited** model only warns, naming both
+    remedies — refusing a bundle over a settings-file key, for a session tldrx does not control,
+    would be asserting more than is known. `--commit` says nothing: that money is already spent.
+
+- **`test/merge-wave.test.ts` no longer fails a wave over another wave's log directory (#95, #97).**
+  `merge-wave.sh` writes its logs to `${TMPDIR:-/tmp}/mw-$$`, and the test that asserts a green run
+  cleans up after itself diffed a listing of the machine's SHARED tmpdir. Every wave on the box
+  writes there — this file's other tests, every sibling agent's wave, and the real merge wave whose
+  `bun test` is running this very file — so the assertion was about ONE run's cleanup and measured
+  the whole machine. It went red twice on trees whose diff touches neither the script nor its tests:
+  **#95**, on a sibling's DELIBERATELY KEPT red-wave log (`mw-35458`, carrying another run's
+  `poison.txt` merge — merge-wave keeps a failed run's logs on purpose, "every FAIL above names the
+  directory it kept"); and **#97**, on a CONCURRENT invocation's LIVE `mw-15412`, twelve minutes into
+  #90's wave, with `main` left at an unpushed merge commit. Both cost a full re-gate, and the same
+  trees passed standalone.
+  - **Every invocation now gets a private `$TMPDIR` inside its own sandbox**, so "did THIS run clean
+    up after itself" has an answer that does not depend on what else the box is doing. Per
+    invocation, not per sandbox: this file deliberately runs two waves at once. The scan is read only
+    after the run has exited, so anything left in that root is a genuine leftover and no liveness
+    rule has to be guessed at — a concurrent wave's live `mw-<pid>` is, correctly, invisible from
+    there.
+  - **What the test asserts about merge-wave is unchanged.** A green run that stops removing its log
+    directory still fails it, and a red run that stops keeping the one it names still fails it — both
+    re-measured by mutating `scripts/merge-wave.sh` and watching the assertions fire.
+  - **A red run's kept logs now leave with the sandbox.** They were being written to the machine's
+    tmpdir and never removed — 1300 `mw-*` directories had accumulated there by the time #95 was
+    filed, including the one it tripped on.
+
+- **A story's first attempt is dispatched at what the plan priced it, not at that figure halved
+  again (#91).** `03-plan/budget.yml` has been read since 2026-08-30, but the price it carries was
+  divided by the worst case ONE story can be asked for — `MAX_ATTEMPTS × (1 + REVIEWER_SHARE)` =
+  2.5 — before a single attempt had run. Measured on run `260901-leaderboard-v2` (finding F-4):
+  Delivery priced S2 at $2.10 of a $3.85 Build stage and the developer was dispatched under
+  **$0.84**. A deliberately-atomic large story starved on the one attempt that mattered while
+  trivial ones carried slack, and the plan's own measured pricing — the whole point of writing the
+  file — was thrown away in the arithmetic that read it.
+  - **`developerPriceDivisor(attempt)`** now decides it. Attempt 1 is the pass Delivery priced and
+    gets `price / (1 + REVIEWER_SHARE)` — the whole price less the reviewer's derived quarter, so
+    S2's ceiling goes $0.84 → **$1.68**. Attempt 2 is a contingency nobody priced and keeps the
+    pre-#91 figure, `price / (MAX_ATTEMPTS × (1 + REVIEWER_SHARE))`. No attempt is ever handed less
+    than it was handed before.
+  - **The even split is untouched.** A plan with no `budget.yml`, one that does not validate, one
+    priced in `host-tokens`, and any story the plan did not name all still get
+    `stage / (stories × attempts × 1.25)`, pinned by tests that were green before this change and
+    are green after it.
+  - **What this costs, stated.** The worst case one PRICED story can be asked for goes from
+    `0.8 × price` to `1.2 × price`. The phase ceiling is metered once, at stage entry —
+    `runNext.runExecutor` skips the brake while a stage is `running` — so nothing re-checks the
+    envelope between two spawns of the same headless `runAll`; that is the window
+    `REVIEWER_FLOOR_USD` already opens by design. `priceScale` still holds the sum of the declared
+    prices inside the stage, and `remainingWork` still clamps the brake's estimate to the stage's
+    own price, so the brake can never refuse more often than it used to. The measured REMAINDER of
+    a story's price would be tighter on the second attempt, and is recoverable (`agent.result`
+    carries `key` = the story id and a row-level `cost_usd`) — it is not read, because the
+    budget-gate hook's `remainingWork` would have to read the same ledger on its hot path to stay
+    in step, and the worst case is `0.8 + 0.4` either way.
+  - **`remainingWork` mirrors the schedule, turn by turn.** `RemainingStory` gains
+    `developerCapsUsd` — one cap per turn still to dispatch, in the order they run — because
+    `cap × turns` stopped being the truth. The turns still to run are the LAST of the story's run
+    of attempts, so a story with one attempt behind it is priced as the attempt 2 it is about to
+    become. `renderRemainingWork` prints `dev $1.20 ×2` while the turns cost the same and
+    `dev $3.60 + $1.80` once they do not, so the total on the line still adds up. A test asserts
+    the mirrored divisor equals the executor's own, as it already did for the three constants.
+
+- **The merge lock now has something to say about raw git in the shared checkout (#89).** The
+  lock serialises merge-wave INVOCATIONS; it never serialised git. Measured 2026-09-02: while
+  agent A's gates were running, agent B typed `git reset --hard origin/main` into the same
+  shared checkout — reflog `reset: moving to origin/main` — and A's merge commit stopped being
+  reachable from `main` mid-gate. #44's gated-HEAD assertion fired for the first time in anger
+  and refused to push. The aftermath was caught; the damage was not prevented.
+  - **`scripts/merge-guard.sh`, installed as a `reference-transaction` hook on every
+    merge-wave run.** That is the only hook git will let abort work in progress: it fires for
+    every reference update and, in its `prepared` state alone, aborts the whole transaction on
+    a non-zero exit. Measured on git 2.50.1 with the guard refusing — `git reset --hard`,
+    `git checkout -B main`, `git merge --no-ff`, `git commit` (including `--no-verify`, which
+    does **not** bypass it) and `git update-ref` all exit 128 with `fatal: ref updates aborted
+    by hook`, and **the ref does not move**. There is no pre-reset, pre-checkout or
+    pre-merge-anything hook; this is the strongest mechanism git offers without a daemon.
+  - **Where the line honestly lies, because it is not where it looks.** `git reset --hard`
+    writes the WORKING TREE before it opens any ref transaction — measured: the file content
+    had already changed to the target commit's while HEAD still pointed at the old one. So the
+    guard saves the **commit**, which is what #89 lost, and cannot save the checked-out files;
+    a wave whose worktree is clobbered mid-gate still gates a tree nobody meant it to, and its
+    gated-HEAD assertion will not fire, because HEAD is exactly where it left it. `git merge`
+    is the better case — refused before the worktree is touched at all. And a hook in a
+    checkout is bypassable by anyone willing to: `git -c core.hooksPath=…`, deleting the hook,
+    or exporting `MW_LOCK_TOKEN`. It is built against accidents, not intent.
+  - **The lock is the sentinel; the marker is what that lock looks like to a human.**
+    `merge-wave.sh` writes `.MERGE-WAVE-IN-PROGRESS` at the root of the shared checkout naming
+    the branch, the owning pid and host, and the way back in. Gitignored, so a wave cannot trip
+    its own dirty-tree guard; removed by `release()` on every path out, INT and TERM traps
+    included. The guard keys on the LOCK rather than on the marker, because only the lock
+    carries an owner and only an owner can be tested for death — a SIGKILLed wave must not
+    leave a file behind that wedges the checkout.
+  - **Scope, so an agent's own work is never in the way.** In the shared checkout every ref
+    update is refused while a foreign wave holds the lock. From a linked worktree only
+    `refs/heads/main` is — refs live in the common dir, so `git update-ref refs/heads/main`
+    typed in a worktree destroys the wave just as thoroughly (measured: it succeeds, unguarded).
+    The holder's own git is recognised by an exported token written inside the lock, so
+    merge-wave cannot deadlock itself, and a lock whose owner is dead is ignored exactly as the
+    waiting loop already ignores it — `scripts/merge-lock.sh` is that one vocabulary, shared by
+    writer and guard so the two cannot drift apart.
+  - **The state name git passes is not portable, and assuming it was shipped a red CI.** macOS
+    git 2.50.1 calls the abortable state `prepared`; the Linux runner's git calls it something
+    else (`fatal: in 'preparing' phase, update aborted by the reference-transaction hook`, run
+    33589554234 on `e0e76a2`). The guard's `*)` arm printed a usage error and exited 2, so on
+    that machine every ref update in every sandbox was refused and the whole merge-wave suite
+    went red at once while macOS stayed green — #49's failure shape exactly. It now recognises
+    `committed` and `aborted` and treats **everything else** as the prepare state, which cannot
+    repeat the blanket refusal: git honours the exit code in the prepare state only, so an
+    unrecognised name is either handled correctly or ignored. The test asks git which states it
+    passes rather than hardcoding them, and additionally probes the names this machine does NOT
+    use — the only way the portability property is testable anywhere.
+
+  - **And the convention it only approximates is now written down**, verbatim and identically,
+    in `CONTRIBUTING.md` and `CLAUDE.md`: agents touch the shared checkout ONLY through
+    `scripts/merge-wave.sh`, every other piece of work happens in their own worktree, and never
+    `git reset --hard` or `git checkout -B main` in the shared checkout. A test asserts both
+    files still say it.
+
+- **A cancelled run is no longer told to retry itself (#86).** `tldrx run cancel` is the way to
+  close a run whose stage FAILED — that is the case `run.cancelled` exists for, and it is a
+  run-level field precisely so the stages keep their failure, "which is history, not state".
+  `deriveRunStatus` read that decision first; `waitingFor` could not read it at all. It is typed
+  against `WaitingRun`, the smallest shape both `RunFile` and the dashboard's tolerantly-read
+  `RunDocument` satisfy, and that shape had `cursor` and `phases` and nothing else — so the answer
+  came from the status of the stage at the cursor, which a cancellation deliberately leaves
+  `failed`. A run somebody had closed reported `waiting.kind: failed` and
+  `"… FAILED — retry: \`tldrx next\`"`, counted as movable, and could wear `← next` on the
+  dashboard. Both screens, because both have read the same derivation since #60.
+  - `WaitingRun` grows an optional `cancelled`, `RunDocument` projects it, and `waitingFor` reads
+    it FIRST — before the cursor checks, in the same order and for the same reason
+    `deriveRunStatus` does. A ninth waiting kind, `cancelled`, joins `WAITING_KINDS`; it is not in
+    `MOVABLE_KINDS`, so `tldrx status` stops offering the run and the dashboard raises no card.
+  - The message names **who closed it, when, and their note** — three facts that were dropped
+    everywhere, on the one screen that exists to say what a run needs. It offers no command:
+    `tldrx next` on a cancelled run already advances nothing, and printing it is what made an
+    operator think there was something left to do.
+
+- **An out-of-order `--commit` no longer leaves a `stage.started` behind (#87).** `runExecutor`
+  stamped the stage `running` and appended `stage.started` on its way in, before the executor had
+  had a chance to say anything. #82 restores the stage's STATUS when that executor turns out to be
+  refusing a sequencing mistake, so `run.yml` comes back byte for byte — but `events.jsonl` is
+  append-only truth and the line could not be unwritten. The live run `260901-leaderboard-v2` left
+  one at 2026-09-02T00:36:37Z: a `stage.started` with no matching `stage.done` or `stage.failed`,
+  which `renderReplay` and anything counting starts read as a start. A `--commit` never STARTS a
+  stage — it settles a cycle a `--prepare` started, and on the ordinary commit the stage is
+  already `running` and no event was emitted anyway — so it no longer emits one at all. The
+  `markRunning` STAMP stays, keeping the original `started_at`, so a `--commit` that does settle a
+  cycle on a stage something demoted still finishes through. The refusal now precedes the event by
+  construction: there is no event. Pinned by measuring the WHOLE log, not a `stage.started` count.
+
+- **An unreadable `result.json` is refused like an absent one, loudly (#88).** Owner decision,
+  2026-09-02. #82 split `PendingError` into `absent` (the host has not written the file) and
+  `unreadable` (the host wrote it and got it wrong), and left the second failing the stage. But
+  nothing was attempted there either — no sub-agent ran, no cent moved, no branch changed — and
+  the fix is the same single command: rewrite the file, run `tldrx next --commit` again. Failing
+  it actively OBSTRUCTED that fix, because it demoted the stage out of `running`, and the
+  phase-budget gate is skipped exactly when a stage is `running` — which is how #82's live run
+  took a `budget.blocked` it had not earned. A host that fat-fingers its JSON paid the same tax.
+  It is #79's model — FORM never costs an attempt, CONTENT/WORK always does — applied to run
+  state.
+  - Corruption does not pass silently. Both call sites in `build.ts` now go through one door that
+    appends a typed **`result.unreadable`** event naming the run-dir-relative path, the parser's
+    own message, the role (`developer` or `reviewer`) and the story, and the refusal line says
+    which file to rewrite and which command to run. It is the only thing a sequencing refusal ever
+    writes; `run.yml` still comes back byte for byte, and `tldrx replay` renders the line.
+  - `parseReview`'s fail-closed rule is untouched and is a different, harsher contract: it governs
+    an envelope that PARSES and is not a valid verdict (unreadable ⇒ `changes`, never `approve`).
+    A file that does not parse at all never reaches it.
+
+- **Watch READS the branch it diffs off `run.yml`; it used to derive one (#90).** The watch stage
+  diffed `feature.epic?.branch` — the `branch:` an epic file DECLARES, written at Plan time before a
+  line of code exists. Under the integration branch model (#57) the Build executor deliberately
+  ignores that value: every epic's stories merge into one `epic/<run-slug>` branch and the epic
+  stays in the plan as a label. So on `260901-leaderboard-v2` the prompt told BOTH watchers,
+  verbatim, that `epic/leaderboard-v2-api` *"does not resolve in scavtopia-workflows. Treat this
+  feature's code as UNSEEN — cite `absent:` rather than guessing at what it emits"* — about a branch
+  nothing had ever cut, while `build.epic_branch` recorded the real one three lines away in the same
+  run. An obedient watcher would have written an all-`absent:` card, and **that card PASSES
+  `claim-sources`**, because an `absent:` citation resolves by construction. Confident, validated,
+  useless coverage; caught only because the host's brief carried the real branch independently.
+  - **The branch now comes from the run's record and nowhere else** (`core/watch/recordedBranch.ts`).
+    `branch_model: integration` → the run's one recorded branch, for every feature. `per-epic`, and
+    every run written before `branch_model` existed → the recorded LIST, with the epic's declaration
+    used *only as a key into it*. A declaration is never returned as a branch: either it is in the
+    record and the record's entry is what is returned, or the answer is "unrecorded".
+  - **A branch the record CLAIMS and the repo cannot find is now a refusal**, not an instruction.
+    It is the run contradicting itself, so the stage refuses before anything spawns (`refused`, so
+    the stage goes back to `ready`), naming the recorded value, the feature and the repo, and saying
+    what fixes it. The treat-as-UNSEEN instruction survives for the one honest case — the record
+    names no branch at all — where the prompt now cites `build.epic_branch` rather than a name it
+    made up. The two absences used to render as the same sentence.
+  - Thirteen tests, red first (6 red / 6 green before, 13 green after): the derivation reproduced
+    under both branch models (the prompt's branch lines asserted byte for byte against
+    `build.epic_branch`), the loud path in both `headless` and `--prepare`, `--commit` deliberately
+    NOT refusing a turn that is already paid for, and the honest absence.
+
+- **A recorded `default_branch` that does not resolve is incoherent state too (#92).** #90 taught
+  Watch to tell a run that recorded NO branch (honest) from a branch its own `build.epic_branch`
+  claims and the repo cannot find (incoherent, refused). It left the third case alone: the BASE of
+  the same diff. `.tldrx/workspace.yml` declares `default_branch: main` for a repo where `main` does
+  not resolve, and the prompt rendered *"`main`, the `default_branch` of api, does not resolve
+  there. Treat this feature's code as UNSEEN — cite `absent:` rather than guessing at what it
+  emits"* — inviting the same all-`absent:` card that passes `claim-sources` and covers nothing.
+  Narrower blast radius than #90 (a misdetected or renamed default branch, not an every-run
+  derivation), same shape: **a value the workspace RECORDS that the repo cannot find.**
+  - **Watch refuses at `--prepare` and headless**, naming the repo, the recorded value and
+    `.tldrx/workspace.yml`, and saying that the value was DETECTED from that repo — so a repo with
+    no such branch means the record has gone stale (a rename, a fresh clone with only remote
+    branches, a misdetection). `epicDiff` carries it as `baseMissing`, the sibling of #90's
+    `branchMissing`, and checks the base FIRST: a base that is not there voids every diff in that
+    repo rather than emptying one, so there is nothing to learn by asking about the branch. On
+    `--commit` it stays tolerant, the same call #90 made — those cards are already paid for.
+  - **`tldrx doctor` reports it**, which is where the issue asked whether it belonged, and the
+    answer is yes: this is a workspace RECORDING something false, and `doctor` is the one command
+    whose job is what this machine and this workspace actually have. One
+    `git rev-parse --verify --quiet` per declared repo, nothing fetched or written. A **warning**
+    that never moves the exit code — `healthy` is about the TOOLS this machine has, and a repo can
+    legitimately be mid-clone or mid-rename on a developer's box; a `doctor` that exits 1 for that
+    is a `doctor` people stop running. A repo that is not on disk, has no `default_branch` recorded,
+    or is not a git repo is reported as SKIPPED with its reason, never folded into "all resolve".
+  - **`boundary` has the gap and deliberately keeps its verdict — measured, not assumed.** It does
+    diff `<default_branch>...<epic_branch>` (`boundary.ts`, `base:` read from `workspace.yml`;
+    changing the record from `main` to `trunk` changes the target's `base` byte for byte), and an
+    unresolvable base came back `{ok: true, detail: "n/a (nothing could be diffed: \`trunk\` does not
+    resolve in app)"}` — the gate passes GREEN and stops measuring for as long as the record is
+    wrong. It is **not** made to refuse: it spawns nothing and writes nothing, so the fault costs it
+    a measurement rather than producing a false one, and `boundary.ts`'s own contract is that it
+    "must not refuse a gate for a reason that has nothing to do with the boundary" — one stale
+    record would otherwise brick every Build gate in the workspace. What changed is the WORDING: its
+    two absences used to be the same sentence (`` `<ref>` does not resolve in <repo> ``), so a
+    merged-and-deleted epic branch (nothing to fix) read exactly like a stale `default_branch`
+    (`n/a` until someone fixes it). The base's reason now names the record and points at
+    `tldrx doctor`; the branch's is untouched.
+  - `.tldrx/workspace.yml` is spelled once, in `paths.ts`, now that three operator-facing messages
+    tell someone to go and edit it.
+  - Fifteen tests, red first (8 red / 7 green before, 15 green after), including the verbatim
+    pre-fix string captured off the real prompt, the `--commit` tolerance, doctor's warning level
+    asserted through `runDoctor` end to end, and the boundary determination pinned on all four
+    counts — that it reads `default_branch`, that it stays `ok`, that its reason now names the
+    record, and that a missing epic branch keeps its own wording.
+
+- **The dashboard says the framework's CURRENT vocabulary, not 0.2.0's.** The live page
+  shipped in 0.2.0 and has had one change since; the framework has had a great many. An audit of
+  `src/core/dashboard/` against today's `run.yml` and `waiting.ts` found seven words the files use
+  and the page dropped. A page that drops a word does not look wrong — it looks finished, which is
+  worse, and every one of these was measured on a rendered fixture before it was fixed.
+  - **`prepared` and `running` rendered as "nothing".** `waitingFor` has had eight kinds since the
+    2026-08-29 audit; the WAITING ON column named five and fell through to `nothing — <status>` for
+    the rest. So a host-attended run with a `--prepare` bundle on disk — the entire attended loop —
+    said it was waiting on nothing while `isMovable` had it wearing `← next` on the same row. Every
+    kind now prints `waiting.message`, which is the sentence the CLI already prints, and `prepared`
+    joins gate/answer/failed as a card: it is a run waiting on a person.
+  - **An `agent` gate was counted as a human one.** The execution-path eyebrow read
+    `N human, M auto`, so a run opened `--gates what:agent,plan:agent,build:agent` reported as
+    all-human — the exact opposite of what it was set up to do, in the one number that eyebrow
+    exists to give. It now counts all three, the same arithmetic `renderGates` does.
+  - **Nothing showed what an agent gate was signed over.** `run.yml` records the verdict, the
+    sample, what resolved and what was refuted, and the run-relative path of the COMMITTED evidence
+    note; the model dropped the whole block, so `agent by reviewer` and `human by alan` read as the
+    same kind of fact. A human signature is a name and a person accountable for it. An agent's is a
+    name and nothing, unless what it checked is beside it. Now `path[].gateEvidence`, printed under
+    the signer — the path as text, never a link.
+  - **`$0.00 spent` on a run whose turns a host session paid for.** `spentUsd` is metered dollars
+    and the meter drew it as the whole story: the exact failure `unmeteredNote` exists to stop the
+    CLI making, made in a progress bar. `attendedBy`, `unmeteredTasks` and `hostTokens` now ride the
+    model, the card prints `attended: host` in the CLI's own words, and the two currencies are shown
+    side by side and never added — there is no exchange rate between a metered dollar and a host token.
+  - **`build.branch_model` (#57) and the epic branches were invisible** in a section headed
+    *Plan & build*. A chained plan on one integration branch and independent per-epic branches drew
+    the identical table. A null model is reported as unrecorded, never guessed at as `per-epic`.
+  - **A `stale` stage looked finished.** A stage left behind by `tldrx reject --stage` wears
+    `done` with its outputs still on disk, derived from a decision that has been withdrawn. It now
+    wears a `stale` chip too.
+  - **The page claimed it read `events.jsonl`.** It never has (measured: nothing in `model.ts` or
+    `loadPhaseArtefacts` opens the ledger). That sentence is why a reader could not tell an empty
+    ledger from an unread one, so the *How to use it* tab now says which files it reads, that the
+    ledger is not among them, and which commands do read it. The Watchers tab, which existed
+    precisely to be honest about a gap, printed an invented card shape whose field names matched
+    none of the seven `Watcher` actually carries; it now prints the real one and names
+    `tldrx watch list`.
+  - **`review` was grey.** A `PLAN_STATUSES` value the page's tone function had never heard of, so
+    a story in review landed in the same colour as one nobody had started.
+  - **The test that should have caught all of this asserted the opposite.** `WAITING_KINDS` is now a
+    value rather than a bare type union, and the chain fixture's coverage test names the three kinds
+    it cannot hold instead of claiming it covers every kind — the claim that let `prepared` go
+    unrendered for months. `modelVersion` stays at `3`: every model change here is an ADDITION, and
+    the rule in `model.ts` is that only a removal or a change of meaning bumps it.
+  - **What the same audit found and did NOT fix is #85**, because each of it is a decision rather
+    than a patch: reading `events.jsonl` (operator notes, story attempts and reopens), reading
+    `budget.yml` (per-phase ceilings, `on_exceed`, whether `next` is affordable), reading the
+    watcher cards, and surfacing a preflight refusal. #86 is a separate bug the audit turned up in
+    `run/waiting.ts` — a cancelled run is offered a retry — which both screens share.
+
+- **The #80 guard's kind sweep could not tell a regex literal in CODE from one quoted in a
+  COMMENT (#83).** The marker half of that guard has a discriminator and says why: a `[` inside a
+  regex opens a character class, so code that matches the `[src:` marker must spell it `\[src`
+  while prose writes `[src: …]` bare. The kind half had no equivalent, because prose quoting a
+  regex LITERAL copies it character for character — there is no signature inside `/^F\d{3,6}$/`
+  that a real production has and a quoted one lacks. #81 paid the tax: the doc comment explaining
+  which two literals had just been DELETED tripped the sweep in the very file that had stopped
+  classifying, and had to be reworded to name the shapes without writing them.
+  - **The discriminator is the line's ROLE, not its characters.** A line that is comment prose —
+    a whole-line `//`, a block opener, or one of its `*` continuation lines — is not a
+    classification, and the sweep now reads the file with those lines removed.
+  - **Line-local, holding no state, and that is the design rather than a shortcut.** #80's own
+    first attempt reached for a lexer and was bitten twice: eleven offenders that were all doc
+    comments, then five more that were template literals nested inside interpolations once
+    comments and strings were stripped. This predicate never looks inside a string, so an
+    unclosed `/**` in a prompt cannot unbalance it and blind the sweep to the code below —
+    asserted, not assumed. What it cannot lex it keeps: a shape quoted in a trailing comment
+    after code is still read as code, so this half still fails CLOSED.
+  - **Measured in both directions on the real tree, not only on fixtures.** Re-injecting #81's
+    exact doc comment into `validateFactsFile.ts` leaves the sweep green; injecting two real
+    productions as CODE into the same file turns it red with `validateFactsFile.ts classifies
+    answer, fact` — the identical message #81 saw for prose. The sweep is also no longer
+    vacuous by assumption: a test pins the kinds the canonical grammar itself trips
+    (`answer`, `cmd`, `fact`) and the six productions the deleted `core/map/srcToken.ts`
+    carried, so a production that stops matching anything becomes visible rather than silent.
+  - **What it removed, measured over the tree.** 28 files matched at least one production
+    before, 20 after; the 8 that stopped matching were doc comments mentioning `→ exit` and
+    nothing else — each one production away from #81's surprise red. No file lost a
+    classification: the most kinds any file decides is still 3, and it is still `srcToken.ts`.
+
+- **A handshake called in the wrong ORDER no longer fails the stage and the run (#82).** On the
+  live run `260901-leaderboard-v2` (2026-09-02T00:36Z) the driver ran `tldrx next --commit
+  --review` while no reviewer bundle was out. The framework said exactly the right thing — "no
+  reviewer bundle is out — run `tldrx next --prepare --review` first" — and then emitted
+  `stage.failed` twice and flipped the run to `status: failed`, so a mistyped command needed an
+  explicit recovery. Nothing had been attempted: no spawn, no cent, no branch, no story moved.
+  Same family as #78/#79 (form vs work), applied to state transitions instead of attempts.
+  - **Five refusals are reclassified**, all of them the handshake asked for by the wrong end:
+    `--commit --review` with no reviewer bundle out; `--commit` with no story `in_progress`;
+    `--prepare --review` over a story whose developer half has not run; and either `--commit`
+    before its `result.json` has been written. Each now exits `1`, names the command that fixes
+    it, and leaves `run.yml` byte for byte as it was — no `stage.failed`, no status moved, no
+    task recorded, no cent metered.
+  - **This is the behaviour the framework already had one layer over.** A single-agent stage sent
+    `--commit` before its `--prepare` has always returned exit `1` and touched nothing
+    (`commitStage`). Build was the outlier because Build owns its own middle, so its refusals came
+    back as `ok: false` and `runNext` could only read that as a failed stage. The fix gives Build
+    the same door rather than inventing a second contract.
+  - **`sequencing` is a second flag, not a widening of `refused`.** `refused` sends the stage back
+    to `ready`, which is right for a precondition an operator must go and fix — a dirty repo, a
+    red base tree — because the cycle cannot continue. A sequencing refusal is a cycle that is
+    perfectly fine, held by the wrong end: the bundle is still out and sending the stage back to
+    `ready` would throw away the state the next command needs. Honoured only when the outcome
+    carries no tasks, no cost and no epic claim, so a refusal that spent something still records
+    it; defaulting to "record it" is the direction a mistake here is recoverable in.
+  - **The bug had a price, and it was the budget brake.** `runExecutor` skips the phase-budget
+    gate exactly when a stage is already `running`, because a Build stage hands out one story per
+    `--prepare`/`--commit` cycle. Failing the stage demoted it out of `running`, so the next
+    `--prepare` was priced as a fresh stage start: on the live run that was a `budget.blocked`
+    ten seconds after the refusal, $2.66 left against a re-charged $4.32 estimate, for money
+    already partly spent. Reproduced in the fixture and pinned.
+  - **`PendingError` carries a typed `kind`** — `absent` (the file was never written; a step of
+    the handshake) vs `unreadable` (somebody wrote it and got it wrong; still a failure). Typed
+    rather than matched on the message, for the reason #79 gave: a caller reading the words breaks
+    the moment the message improves.
+  - Genuine failures are untouched and pinned: a red Definition of Done still blocks its story, a
+    reviewer that dies is still recorded as a failed check, and a plan that cannot be loaded is
+    still `stage.failed` with a `failed` run. Two existing pins moved deliberately, both of them
+    asserting the old exit `5` of a pure sequencing refusal.
+
+- **`facts.yml` ids and citable ids were two spellings of one shape; now they are one constant
+  (#81).** `src/core/facts/validateFactsFile.ts` defined its own `F` and `Q` id patterns, agreeing
+  with `SRC_PATTERNS.fact` / `.answer` character for character, with nothing asserting that they
+  must. Found while fixing #80 and filed rather than fixed there, because it is not #80's defect:
+  the `[src: …]` token grammar has exactly one reader now, and what was duplicated here is the id
+  SHAPE that grammar happens to share with the facts file, reached by a different reader answering
+  a different question — which is why #80's kind sweep allowlisted the file rather than folding it in.
+  - **They must agree, and the DIRECTION is what makes it load-bearing.** One string makes one
+    trip: `formatFactId` mints `F102`, `validateFactsFile` admits it into `.tldrx/memory/facts.yml`,
+    and `classifySrc` reads that same `F102` back out of a `[src: F102]` token for `knowledgeFile`
+    to resolve against the store. So every id the facts file accepts must be citable, or a fact
+    exists that cannot be cited and is invisible to every reader downstream of it. Nothing in that
+    pipeline ever runs both readers on the same string, so the drift would have been silent. (The
+    other direction is merely untidy: a citable id no facts file holds is caught later, as a
+    citation that resolves to nothing.)
+  - **Derived, not merely asserted.** `validateFactsFile.ts` imports `SRC_PATTERNS.fact` and
+    `.answer` — the same publication surface `srcGrammarContract.ts` already generates the
+    documented grammar from — so there is one spelling rather than two that match. The #80
+    allowlist entry is deleted along with the need for it, and that sweep now covers the file like
+    any other; an exemption that has stopped being needed is one that will some day cover something
+    it was never granted for.
+  - **A behavioural guard too, because sharing a constant only holds while nobody re-types a
+    literal.** The kind sweep needs TWO productions in one file before it fires, so a single
+    respelled `F` shape slips straight past it — measured by sabotage, not assumed.
+    `test/map-citations.test.ts` now runs both readers over a table of ids spanning both digit
+    boundaries and compares their verdicts; loosening either shape by one digit turns it red with
+    `F12: facts.yml holds it but the grammar refuses it`.
+  - **One drift had already happened, in prose.** The `source.q` refusal read `expected ^Q\d+$ or
+    null` while the reader ran `^Q\d{1,6}$` — the shape's third spelling, telling an author a
+    seven-digit question id was acceptable and then refusing it. Both refusal messages are now
+    generated from the pattern they enforce, so neither can drift from its reader again.
+- **The docs site said 0.3.1, named a flag that does not exist, and told readers to hand-write
+  a mandate the CLI now prints.** A cold read of every page against `tldrx <cmd> --help`, EN and
+  ES. Three claims were false rather than merely stale: the version, printed on six pages, was two
+  releases behind what `npm i -g tldr-experts` installs; `--max-budget-usd` was named as a `tldrx
+  next` flag in both budget pages, but it is what tldrx passes down to `claude` and `tldrx next
+  --max-budget-usd 3` exits `1` on an unknown flag; and `guides/driving` said "there is no keyword
+  for the third part — the mandate is prose you write" three weeks after `tldrx drive
+  --attended|--unattended` shipped to print exactly that, versioned with the package. Also
+  corrected: the hero claimed a gate you own at the end of *each* stage, where `feature` ships two
+  `auto` ones; the quickstart said "five role experts" one line under output reading `experts 6
+  seeded`, and reported `$0.31 of $4.00` on a run opened at `$5.00` without saying the second
+  figure is the stage's ceiling.
+- **The site covered none of the last release's commands, and the Spanish mirror had fallen a
+  release behind the English one.** Added across both languages: `story reopen --for-fix`,
+  `questions lint`, `retro --all --json`, `watch check --execute`, `update_check: off`, `seed
+  answer`, exit `130`, `budget.yml`'s `ceiling_host_tokens` / `on_host_tokens_exceed` (the host
+  economy was documented as unboundable), the integration-branch model for chained epics (#57),
+  a FAQ entry for `tldrx update` and the version notice, and a link to `CONTRIBUTING.md`. The ES
+  CLI reference was additionally missing `update`, `drive`, `retro --all` and `plan schema`
+  outright. `reference/cli` no longer pins a version number that goes stale every release. The
+  Spanish pages' code comments were half-translated; terminal output stays English, the docs
+  author's own `#` comments do not.
+
+- **`tldrx map --check` ran a SECOND, divergent `[src: …]` grammar; there is one grammar now
+  (#80).** `src/core/map/srcToken.ts` was not a thin wrapper over the claim-sources reader — it was
+  a parallel set of regexes, and they disagreed on five axes: its `file` pattern
+  (`[^\s:]+`) refused a path containing a colon, its answer pattern (`^Q\d+$`) had no digit cap,
+  its token pattern was **global** so a citation written mid-sentence counted as one, it stripped no
+  trailing backtick or full stop before matching, and it had never heard of `aidlc:`. So a citation
+  `claim-sources` accepted could be reported by `map --check` as a problem, and one it refused could
+  be reported as fine — the "two readers of one question drift, and the looser one wins the argument
+  at the wrong moment" hazard `core/text/handoff.ts` is written against. It also carried #77's
+  defect on its own path, printing the symptoms `bullet has no [src: …] token` and
+  `unparseable src token` and nothing else.
+  - **The parser is deleted, not wrapped.** `checkCitations` calls `parseSrcToken`,
+    `diagnoseSrcToken` and `describeSrcFailure` from `src/core/text/srcToken.ts`, so every `map
+    --check` grammar failure gets the #77 treatment by construction: the rule id, the rule in its
+    own words, the line as written, and a line that would have passed. `CitationProblem` gained a
+    `rule` field, so a caller can tell "you wrote it wrong" from "it has drifted" without parsing
+    prose. `srcToken(srcs)` — the BUILDER, which had no equivalent on the canonical side — moved
+    across rather than went, and now joins on the same `SRC_SEPARATOR` the reader splits on.
+    `isBullet` stayed behind in `checkCitations.ts`: it is about DOCUMENTS, and mixing the two is
+    what the deleted file got wrong.
+  - **The migration hazard was measured before the switch, not asserted after it.** Both grammars
+    were run over every `.tldrx/map/**` document and init handoff in the two real workspaces
+    available — **39 documents, 692 lines, 435 of them carrying a citation** — and they disagreed on
+    **zero** lines: no multi-token line, no non-bullet citation, and not one token that failed to end
+    its line. The axes where they *would* differ are pinned as a table in
+    `test/map-citations.test.ts`. Unification is **stricter** on six shapes (a mid-line token, a
+    `..` in a path, a 7-digit `Q`, a 4-digit exit code, a backwards line range, a line number of
+    zero) and **looser** on five that the old reader wrongly refused (a path containing a colon, an
+    `aidlc:…#Q<n>` src, a `cmd` whose command holds a backtick, an `absent:` path with a space, a
+    token wrapped in backticks). An `aidlc:` citation used to be misread as repo `aidlc` and
+    reported as an unknown repo; it is a kind now.
+  - **A guard makes re-adding a copy go red** (#48's lesson: deleting one file does not stop a
+    second). Two halves, both on the SHAPE rather than the file name — no file outside the canonical
+    grammar and two documented non-parsers may hold a regex that matches the `[src:` marker, and no
+    file outside it may classify two or more src KINDS. The guard is proven against the deleted
+    file's own two regexes rather than assumed. It also turned up a real second copy of the fact and
+    question id shapes in `validateFactsFile.ts`, filed as #81 rather than fixed here.
+  - **`map --check`'s summary counts the two failures separately.** A citation that does not parse
+    never reached the filesystem, so folding it into "N of M citations do not resolve" reported a
+    denominator it was never in — and, when every problem was a grammar one, printed "3 of 0".
+
+- **A `[src: …]` rejection now states the RULE it enforced, and the grammar is published where the
+  writers read it (#77).** Run `260830-ordering-inventory` lost **three story attempts** to one
+  message. Three review envelopes were refused with "no `[src: …]`" — the SYMPTOM — while all three
+  carried a citation. The host guessed the grammar twice, got it wrong twice, and finally opened
+  `dist/tldrx.js` to extract three rules that were written down nowhere a writer could read them:
+  the token must END its line (`TRAILING_TOKEN_RE` is anchored), a `]` **inside** the token
+  truncates the match (`[^\]]*`), and a `cmd` source needs the real `→`, never ASCII `->`.
+  - **Every rejection on the path names its rule, quotes the line, and shows a corrected one.**
+    `SRC_RULES` (`src/core/text/srcToken.ts`) is fifteen rules, each with the pattern that enforces
+    it and a worked `bad`/`good` pair; `diagnoseSrcToken` maps a failure onto exactly one. The hook
+    denies, the gate's `claim-sources` detail, `parseFixFindings`' `refuted` refusal and every
+    `classifySrc` error now carry the id. The deny messages take the document text and quote the
+    offending line back — naming `L14` and stopping there is what sent the host looking.
+  - **The grammar is GENERATED and spliced into the prompt** (`renderSrcGrammarContract`), following
+    the #35 precedent: kinds from `SRC_KINDS`, patterns from `SRC_PATTERNS` printed `.source` and
+    all, rules and examples from `SRC_RULES`, the four-section rules from the same constants
+    `handoff.ts` enforces. It reaches the writers through the check-contracts registry (every stage
+    declaring `claim-sources` over a `.md` output), the Watch executor — which read `stage.md`
+    raw and so was the one writer never given any contract — and the reviewer prompt, whose
+    `refuted` verdict is held to this grammar and was never told what it is.
+  - **The trap is behavioural, not textual.** `test/src-grammar.test.ts` pushes every documented
+    `bad` back through `diagnoseSrcToken` and every `good` through `parseSrcToken`. Loosen a regex
+    without updating its rule and the suite goes red; a doc that has stopped being true cannot ship
+    quietly. It caught one bug on its first run: `RegExp.source` re-escapes non-ASCII, so
+    `CMD_RE.source` spells the arrow as a `\u`-escape — the contract would have documented "use the
+    real `→`" with the arrow itself written as an escape sequence. `readableSource` decodes it.
+  - Not changed: what the parser ACCEPTS. Tolerating `->` is a product decision and stays open
+    as #77's item 3.
+- **A grammar-rejected review envelope no longer consumes a story ATTEMPT (#78).** Measured on run
+  `260830-ordering-inventory` (2026-09-01): stories S2, S3 and S5 each recorded
+  `check: review · verdict: changes · attempt: 1` over a summary beginning *"I would sign this:
+  every named acceptance criterion is met"*. Those were `fixlist` envelopes refused by the
+  **claim-sources** check — a `refuted` finding whose `[src: …]` sat mid-line, where §2.8's
+  end-anchored parser cannot see it — and each was charged to the story as a failure of its WORK.
+  Three of the run's attempts went on formatting. A malformed envelope is a fault in the reviewer's
+  *report*; conflating the instrument with the result is what this fixes.
+  - **The framework asks again, for free.** A refusal the claim-sources check raised re-prompts the
+    same reviewer for a corrected envelope, carrying what was refused verbatim under
+    `## Your previous envelope was REFUSED`. **Bounded at two** per envelope round (owner decision,
+    2026-09-01): the third refusal is recorded as the ordinary `changes` and costs the attempt, so a
+    reviewer that cannot write the grammar at all still settles instead of looping free. The bound
+    resets when a verdict is finally counted — it is per envelope round, not per story.
+  - **Both doors, one rule.** A spawned reviewer re-prompts itself in-process;
+    `tldrx next --commit --review` leaves the bundle out with the refusal spliced into its
+    `prompt.md`, bins the refused `result.json` and settles nothing. Attempt accounting must not
+    depend on which door a verdict came through, so both go through the same predicate.
+  - **Auditable, because it is bookkeeping.** Each free round appends one `story.review_retried`
+    (§2.9) carrying the story, the attempt it did **not** spend, which retry it was, the bound, and
+    the refusal. `readReviewLedger` counts them, which is how the bound survives a fresh
+    `tldrx next` and the one-envelope-per-process host handshake alike. Each re-prompt is a real
+    metered turn and gets its own task row: it costs the story no attempt, never no money.
+  - **Scope guard, pinned by test.** Only the claim-sources grammar — *widened to every
+    envelope-FORMAT refusal by #79 below, in the same release.* A verdict's CONTENT and a red DoD
+    keep exactly the cost they had.
+  - **The message is #77's, inherited rather than copied.** The re-prompt carries the refusals
+    verbatim and points at the `Citation grammar` section #77 splices into the
+    same prompt, so the reviewer is told which rule it broke, on which line, with a corrected
+    example — and there is no second copy of the grammar to keep in step. The classifier
+    is a typed INDEX over `problems`, not a second list, precisely so #78 could not end up
+    string-matching the text #77 was rewriting.
+
+- **Every envelope-FORMAT refusal gets that same free round, not just the citation (#79).** #78 drew
+  its scope at the claim-sources check because that is what the evidence named, and filed the rest.
+  But `parseFixFindings` refuses an envelope for five other shapes — a `fixlist` that is missing,
+  not an array or empty; a row that is not an object; a row with no `finding` text; a row with no
+  valid `disposition` — and `parseReview` refuses a sixth, a verdict WORD outside the enum (#36).
+  Every one of them is a fault in how the reviewer wrote its *report*, exactly the argument #78 made,
+  and every one of them still cost the story an attempt. Owner decision (2026-09-01, on the issue):
+  all of it, **one mental model — FORM never costs an attempt, CONTENT/WORK always does.**
+  - **Nothing about the mechanism changed.** Same bound (two per envelope round, the third is the
+    ordinary `changes`), same counter, same `story.review_retried` event, same two doors, same
+    per-turn metering. #79 widened *what earns a correction* and touched nothing else.
+  - **The verdict WORD (#36) is now free too, and #36's message is unchanged.** A reviewer that
+    writes `sign` — the gate vocabulary — said nothing wrong about the diff; it reached for a word
+    the story enum does not have. It is told so, by name, and asked again. `Review.formatProblems`
+    is the union of the fix-list index and that verdict fault, because it is also what the corrected
+    envelope's prompt is rendered from: a refusal missing from it is one the reviewer is never told
+    about.
+  - **The guard is the INDEX, which is what survives the next widening.** `ParsedFixlist.format`
+    stays a typed subset of `problems`, built one push site at a time through a single
+    `refuseFormat` helper — never a second list of strings, and never "everything `parseFixFindings`
+    said". The free round is granted only when the index claims **every** reason the envelope was
+    refused, so a refusal about the WORK added later costs the attempt until somebody deliberately
+    indexes it as form. Defaulting to *costs* is the direction a mistake is recoverable in.
+  - **A non-empty `findings[]` is deliberately NOT the content signal**, though it is the obvious
+    candidate. Measured across the nine `aparece-v2` runs (2026-09-01): all 25 recorded review logs
+    carry a non-empty `findings[]` and all 25 are `approve` — the one verdict whose own prompt line
+    says *"Empty on `approve`"*. Reviewers use it as a narrative evidence log whatever the verdict,
+    so gating on it would have made the free round almost never fire and quietly narrowed #78 as
+    well. A judgement about the work is caught where it is actually stated: a declared `changes`
+    raises no format refusal at all, so it costs its attempt.
+  - **Nine tests red first**, each newly-free class proven to re-prompt without spending an attempt
+    and to record the event, plus the bound re-proven on a shape refusal. #78's CONTENT and DoD pins
+    are unchanged and still green.
+
+- **`merge-wave.sh` no longer leaves a conflicted tree behind, wedging every queued sibling (#76).**
+  On a merge conflict the script exited `2` **without** `git merge --abort`, and the `EXIT` trap
+  then released the lock. The conflicted index survived that handover, so the next queued
+  invocation acquired the lock, failed its dirty-tree guard and exited `1` `FAIL dirty tree` having
+  merged nothing — and so did every one after it, until a human ran the abort by hand. Observed
+  live 2026-09-01 by two agents: a sibling's abandoned merge left `UU CHANGELOG.md` plus 14 staged
+  paths in the shared checkout, and cluster L's first merge-wave returned `1` with nothing merged.
+  Under the concurrent multi-cluster pattern, one conflict wedged every other cluster.
+  - **Collect, then abort** — the order `mergeNoFf` already uses one directory over
+    (`src/core/build/git.ts:314-326`). The agent still learns exactly which files conflicted and
+    still has to rebase and retry; the checkout it hands back is the one it was given. This is the
+    same class of hazard the lock was written for: state from one invocation leaking into the next.
+  - **The refusal now names what is dirty.** `FAIL dirty tree` alone accused the caller of leaving
+    junk in their own checkout when, inside the lock, the likeliest cause is another run's residue
+    in the shared one. It now says so and lists the paths.
+  - Proved by a repro that runs the real script against a real conflicting merge and asserts
+    `git status --porcelain` is empty afterwards, `MERGE_HEAD` is gone, `HEAD` has not moved, the
+    lock is released, and a second invocation merges instead of being refused.
+
+- **The built-CLI dashboard test no longer gates a stale `dist/` (#73).** `beforeAll` read
+  `if (existsSync(DIST)) return`, so the tests ran against whatever `dist/tldrx.js` happened to be
+  lying around — locally, a build from before the working-tree changes the run was checking.
+  Measured 2026-09-01 while fixing #60: the guard served a binary built at 14:43, and the model
+  version assertion read `Expected: 3, Received: 2`. That is the lucky direction; the same staleness
+  hiding a regression is silent, and `bun run build` is a separate later step in both
+  `scripts/merge-wave.sh` and CI. CI was safe only by accident — a fresh checkout has no `dist/` —
+  which means the guard only ever applied where it did harm.
+  - **Always rebuild, and the number is the reason.** `bun scripts/build.ts` costs 199 ms cold and
+    55–59 ms warm on the reference machine (3 runs, 2026-09-01) against a ~420 s suite: under 0.05%.
+    A stamp of `src/` would cost more to keep honest than it saves, and `test/build.test.ts:78`
+    already built unconditionally — this file was the exception.
+  - A new assertion pins the property rather than the mechanism: `dist/tldrx.js` may not predate the
+    newest file under `src/`, `scripts/build.ts` or `package.json`.
+
+- **`tldrx drive` fills the mandate's `<run>` in (#75).** The mandate's every command read
+  `tldrx next --prepare <run>`, and the header told the reader to find-replace — 7 occurrences
+  unattended, 5 attended (measured; the issue estimated ~8), by hand, at the exact moment somebody
+  is trying to start a run. One occurrence missed sends a session at the wrong run.
+  - **An id, or the one open run.** `tldrx drive --unattended <run>` or `--run <id>` (the positional
+    wins, `ship`'s order) substitutes **textually and never validates** — an id naming no run is the
+    operator's typo to notice, and the command stays the one thing in the CLI that runs anywhere.
+    With no id, the ONE open run of the current workspace is used.
+  - **It refuses to guess between two.** Where `RunStore.resolve` would call it ambiguous, drive
+    declines to substitute, leaves `<run>` standing and names the ids on stderr — a mandate silently
+    aimed at the wrong run is the bug being fixed, not a smaller version of it. No workspace, no
+    runs, an unreadable `tldrx-work/`: all keep the placeholder and still exit `0`.
+
+- **`tldrx status` no longer calls a RUNNING run "cannot start yet" (#60).** Verbatim from
+  aparece-v2, 2026-09-01: `run 260830-ordering-inventory (…) cannot start yet — it was proposed to
+  follow money-and-payments` / `at 04-build / build · run status running · waiting: prepared` /
+  `blocked by money-and-payments — it is pending`. The run was building, with S1 verified minutes
+  earlier. `triage.depends_on` is an order a split PROPOSED before either run existed, and it was
+  out-ranking what the run was observably doing.
+  - **Observed state outranks proposed order.** One rule, in the resolver both screens read
+    (`src/core/run/dependencies.ts`): a run that has left `pending` has started, and a proposal
+    cannot un-start it. `hasStarted` is the one definition — `pending` is the only status a run that
+    has never run a stage can wear, because runs are created with every stage `pending` and every
+    path to any other value goes through a stage that was `running`.
+  - **The proposal becomes a footnote, not the headline.** A started run renders its own cursor and
+    waiting kind, gains back the command it was denied, and carries
+    `proposed to follow <run> — started anyway` as a secondary line. `blocked by` is now reserved
+    for a run that really cannot move — one that has not started. A run that has NOT started is
+    unchanged: same words, same withheld command.
+  - **It gets the `← next` slot back.** The ordering hint had demoted the only run with work in
+    flight and pointed the owner at the sibling that had not begun. `runnable` is now
+    `movable && (started || nothing outstanding)`.
+  - **`prepared` and `running` get their own summary lines.** Both used to fall through to
+    "is blocked at <phase>/<stage>" — the same wrong word, one layer down.
+  - **The dashboard was making the identical claim** about the same shape and is fixed with it:
+    fixture run `charlie` is `awaiting_gate` behind a `pending` sibling, and the page said
+    "blocked by bravo" while suppressing the gate alert for a signature a person could give right
+    then. `DASHBOARD_MODEL_VERSION` → **3**: `blockedBy` is unchanged and still records the
+    proposal, the new `runs[].started` says whether it still holds anything back, and `runnable`
+    reads `true` for that one shape where it read `false`.
+
+- **A cost/token declaration on a story commit now attaches to the build task (#68).** Measured on
+  two live runs: the leaderboard host declared $2.25 on a story commit and `tldrx cost` kept
+  04-build at $0.00 — "the declaration didn't attach to the build task the way it did for
+  what/how/plan" — and ordering 260830 showed the same shape. The run's recorded build spend was a
+  floor, budget arithmetic ran blind on the most expensive phase, and ~8M host tokens were invisible.
+  - **The seam, not the accounting.** `ExecutorContext` never carried `--cost-usd`/`--tokens`, so
+    Build's `commit()` could only read the envelope's own `cost_usd` and wrote
+    `round2(result.cost_usd ?? 0)` — a METERED `$0.00`. It now resolves
+    `options.costUsd ?? result.cost_usd`, which is exactly what `commitStage` does for what/how/plan,
+    and the declared tokens ride onto the task row and the `agent.result` payload the cost report
+    reads. The reviewer half (`--commit --review`) takes the same precedence.
+  - **Nothing declared is UNMETERED, not $0.00.** `cost_usd: null` + `metered: false`, the spelling
+    every other host turn already gets, because `$0.00` is a measurement and a false one.
+  - **Backward-safe.** A declaration is never allowed to overwrite a measurement — an envelope that
+    reports its own cost still wins over nothing declared — and no recorded zero is rewritten. Only
+    new declarations attach.
+
+- **`onStderrLine` on the runtime seam, so progress prints before the verdict it produced (#67).**
+  `SpawnOptions` had `onStdoutLine` and nothing for the other stream, so stderr — where every tldrx
+  UI writes its progress — could only be handed back as one string at exit. `tldrx learn` printed
+  `01-what/what done — $0.31 of $4.00` and only THEN the `[00:00] writing …` lines that produced it.
+  - Implemented in **both** `bunRuntime.ts` and `nodeRuntime.ts`, with the same contract
+    `onStdoutLine` already has: the full text still accumulates in `SpawnResult.stderr`, a trailing
+    partial line is delivered at close, and omitting the callback leaves the buffered path
+    byte-for-byte as it was. Node gets its own `LineSplitter` per stream — one shared buffer would
+    splice a half-written stdout line onto the front of a stderr line.
+  - **The learn workaround is deleted**, per the instructions it carried since phase 1:
+    `realStepRunner` passes the callback and `playChapter` no longer writes `result.stderr` after the
+    step. Measured by playing chapter 2 by hand on both checkouts: the summary was at line 107 with
+    its progress lines at 109-124, and is now at 123 with them at 105-121.
+  - The tests turn on TIMING, not content: a buffered implementation could hand the same lines to the
+    same callback by splitting at exit, so the line is required to arrive while the spawn's promise
+    is still pending.
+
+- **`tldrx plan` has a section in the CLI reference (#55).** `docs/guide/08-cli-reference.md` had no
+  `## tldrx plan` heading at all, so `sync-dod` — the one mechanical repair for stories whose dod
+  block a `workspace.yml` edit orphaned — was undocumented. The section states the four per-line
+  outcomes, the git-history ancestry, the `.md.bak`, and that it runs no agent and moves no cursor.
+  The #54 docs-coverage test is widened from `run` to `run` and `plan`, and now also asserts the
+  heading rather than a passing mention. It is still a scoped list rather than a generalisation over
+  every command: `hook`'s seven scripts are a deliberate spelling, and `note` and `ship` have the
+  same gap #55 was about (filed separately).
+
+- **`tldrx update`, and a one-line notice when a newer version exists (#62, owner decision
+  2026-09-01: on by default, with an opt-out).** The owner installed 0.4.0 on a second machine and
+  found there was no way to ask the tool to update itself, and no way to be told a newer one existed.
+  - **`tldrx update` is `npm i -g tldr-experts@latest`, run for you**, plus the part a wrapper does
+    not give you: the CHANGELOG between the version you had and the version you now have.
+  - **The new version is READ BACK from what npm installed** —
+    `$(npm root -g)/tldr-experts/package.json`, with the delta taken from the `CHANGELOG.md` beside
+    it. The process printing that line is the OLD build and cannot know what the new one is; when the
+    read-back fails it says so and prints no changelog rather than inventing one.
+  - **The notice never touches the network on the hot path.** A command reads
+    `~/.tldrx/version-check.json` and nothing else. The registry call happens in a DETACHED child
+    (`stdio: "ignore"`, `unref()`) spawned after the output is written, and its answer is for the
+    NEXT invocation. Cached for 24 h; silent on any network failure, on a body that is not the JSON
+    it asked for, and on a home directory it cannot write.
+  - **Never in `--json`, never during a hook, never off a terminal.** `tldrx hook` and
+    `tldrx statusline` are suppressed by name (spec §0: a hook is deterministic), `--json` anywhere
+    in argv is suppressed, a non-TTY stdout is suppressed, and so is CI. The line itself goes to
+    stderr, so no command's stdout changes shape.
+  - One line, in the issue's own wording: `tldr-experts 0.5.0 available (you have 0.4.0) — tldrx
+    update`. Asserted as an exact string, because "roughly this sentence" is how one line becomes
+    three.
+  - **Opt out** with `TLDRX_UPDATE_CHECK=off` for a shell (spelled like `TLDRX_UI` and
+    `TLDRX_CLAUDE_BIN`; `0`, `false`, `no` and `never` also work), or `update_check: off` in
+    `~/.tldrx/config.yml` for the machine. A config file that does not parse is not an opt-out and
+    not a crash.
+
+- **`budget.yml` no longer adds host tokens to dollars (#61, owner decision 2026-09-01).**
+  `validateRunBudget` summed EVERY phase ceiling and compared the total to `ceiling_usd` — but since
+  the `economy:` label landed, a phase priced in `host-tokens` carries a host-session token allowance
+  and not money. A realistic allowance therefore made a valid file invalid
+  (`phase ceilings sum to 200018 > ceiling_usd 25`), and it did so in the one check that fails
+  **closed**: `RunStore.open` threw, and the budget-gate hook then denied every spawn on the run.
+  - **Separate ceilings, per economy.** `ceiling_usd` stays dollars-only and a new optional
+    `ceiling_host_tokens` — at the run level and per phase — is the host-token allowance. The two are
+    never added and never converted: there is no exchange rate here, and inventing one would be a
+    guess about a price, which is the whole reason the label exists.
+  - **The sum runs once per economy.** Σ the `metered-usd` phases against `ceiling_usd`; Σ the
+    `host-tokens` phases against `ceiling_host_tokens`. The dollar half is byte-identical arithmetic
+    with the same refusal in the same words.
+  - **Additive, and absence changes nothing.** With no `ceiling_host_tokens` declared, the token sum
+    has nothing to compare against and is not checked — deliberately the lax side, since the only
+    other number on the run is dollars. The compat bar is the live `260830-ordering-inventory`
+    budget.yml, mid-Build in another workspace while this was written and asserted verbatim in
+    `test/economy.test.ts`: it validates, its phase sum of exactly 62.00 against a 62.00 ceiling still
+    passes, one cent more is still refused in the same words, and every phase still reads
+    `metered-usd` with no token ceiling.
+  - `hostTokenCeiling` (f353d8d, #22b) prefers the new field and **still falls back to `ceiling_usd`**,
+    because the files written before it existed put the allowance in that one unlabelled scalar and
+    they are still on disk. The emitter round-trips both, so `budget raise` cannot erase a token
+    ceiling it rewrites past.
+  - The workaround the issue documented is gone: `test/hooks.test.ts` and `test/economy.test.ts` no
+    longer raise the run ceiling to `60000`/`100000` — dollar figures that meant nothing — to let a
+    token phase validate.
+  - **`RunStore.ceilingsToWrite` now carries a phase's `economy` and `ceiling_host_tokens` from disk**
+    beside its `ceiling_usd`. That seam re-reads CEILINGS before every save so a raise landing
+    mid-stage is not clobbered by the copy the process opened with — and a phase's ceiling is three
+    fields, not one. Preserving the number without the label would have been the worse half of both:
+    a token allowance kept on a phase this reader had been told is priced in dollars, where
+    `hostTokenCeiling` can no longer see it. For a file with no labels and no token ceilings — every
+    budget.yml on disk today, the live one included — all three are identical on both sides and the
+    save writes exactly what it always wrote.
+
+### Removed
+
+- **`templates/story.md` and `templates/epic.md` are deleted (#48, owner decision 2026-09-01,
+  option (a)).** They stated the Plan front-matter schema, shipped in the npm package, and
+  `grep -rn 'story\.md' src/` found nothing that read either one. Since 3ae0ce9 the live copy is
+  generated: `src/core/plan/schemaContract.ts` builds the story, the epic and `waves.yml` from
+  `STORY_KEYS` / `EPIC_KEYS` / `PLAN_STATUSES` / the `MAX_*` constants and splices them into the Plan
+  prompt. The drift guard 7ac298c added held the two files to that contract; deleting them removes
+  the second copy instead of maintaining it.
+  - **The consumers were tests, and they now generate.** `test/plan.test.ts` and
+    `test/plan-schema-contract.test.ts` were the only readers; both take the story and the epic from
+    `planContractExamples()`. Nothing in `src/`, `stages/`, `workflows/`, `plugin/` or `docs/` read
+    either file, and `run new` never copied them.
+  - **`templates` stays in `package.json` → `files`.** The directory still ships eleven templates
+    that ARE read at runtime — `templates/expert.md` and `templates/experts/<role>.md` are read by
+    `createExpert.ts` and `roleExperts.ts` in an installed package — so removing the entry would
+    break `tldrx expert create` to delete two files that no longer exist.
+  - **The drift guard changed meaning and kept its teeth.** It now asserts the GENERATED story and
+    epic validate through `validateStoryFile` / `validateEpicFile` with keys equal to `STORY_KEYS` /
+    `EPIC_KEYS` in order, that neither file is back on disk, and that no OTHER shipped template has
+    grown the same front matter under a new name. Proven, not assumed: setting the example's
+    `status:` to `wip` turns 4 tests red, and renaming `test_plan` in the generator turns 6 red.
+
+## 0.4.0 — 2026-09-01
+
+### Changed
+
+- **Dependent epics share ONE integration branch (#57, owner decision 2026-09-01, option (a)).**
+  One-branch-per-epic assumes the epics are independent. Run `260829-scoring-leaderboard` planned
+  E2 (the API) with E3 and E4 (mobile) consuming it, and a downstream story's base — cut from its
+  own epic branch, itself cut from `main` — could not see the upstream epic's merged work. It broke
+  twice, and both times the host fast-forwarded the EPIC branches by hand: cross-epic surgery done
+  with a feature built for stale STORY bases (design §F.2), which collapses branches the owner may
+  have meant to merge separately. A run whose epics form a dependency chain now cuts a single
+  integration branch and the epics become labels.
+  - **Detected at PLAN time, from what the plan already says.** A story whose `depends_on` names a
+    story in another epic IS the chain. `validatePlan` reports the cross-epic edges it read
+    (`PlanReport.epicChain`, deduplicated per epic pair), and the `plan` gate check states the branch
+    model in its passing detail — `epics form a chain (E3→E2, E4→E2) → single integration branch
+    \`epic/<run-id>\`` or `independent epics → one branch each`. The owner never discovers it
+    mid-Build, which is the half of #57 that was not about branches at all.
+  - **`epic/<run-id>`, not a new namespace.** `EPIC_BRANCH_RE`, `watch`'s feature-slug extraction,
+    `ship`, `boundary` and the `--reuse-epic` guard are all keyed on the `epic/` prefix; an
+    `integration/…` branch would have changed every one of them to buy a word. The run id IS in the
+    name — unlike an ordinary epic branch, which is deliberately unscoped because an epic is the unit
+    a team merges — because an integration branch belongs to one run by definition.
+  - **One run-scoped epic worktree**, `_epic-<run>-integration`, because git will not check one branch
+    out in two worktrees. It is picked up by the `_epic-<run>-` prefix that the §2.8 src resolver and
+    the run-close cleanup (#16) already enumerate, so the lifetime decision from that change holds.
+  - **Backward-safe by an ABSENT key.** The Build executor records what it used in `run.yml`
+    (`build.branch_model: per-epic | integration`, additive and optional). A run.yml that names
+    branches and no model predates the key — the three closed runs, and any run mid-flight — and stays
+    `per-epic`, so it resumes on the branches it already cut rather than being re-pointed at one that
+    was never cut. A model, once written, is never rewritten.
+  - **Nothing else moved.** The dirty-tree refusal, the foreign-epic refusal and `--reuse-epic`, the
+    gated HEAD, `git merge --no-ff` into the epic worktree, and the story-base fast-forward all behave
+    exactly as before; only WHICH branch they name changes. `tldrx ship` needed no change: a chained
+    run claims one branch, so it opens one PR with no `--branch`, and an unchained multi-epic run still
+    asks which.
+  - The acceptance test is the leaderboard shape, passing: S1 in E1 writes a file, S2 in E2 depends on
+    it, and S2's worktree holds that file with **no `story.base_fastforwarded` event** — the base was
+    right when the developer was dispatched, not repaired afterwards. Red before the change (`Expected:
+    true, Received: false`).
+
+- **An epic worktree now lives for the RUN's lifetime, not the Build stage's (#16, owner decision
+  2026-09-01, option (a)).** The shipped half of #16 made a `file` src resolve against
+  `.tldrx/worktrees/<repo>/_epic-<run>-<epic>` before the working tree — and then `BuildSession.finish()`
+  removed that directory before the Build handoff was even written, so unless the operator had typed
+  `--keep-worktrees` a later Watch stage had nothing to resolve against and the fix bit only under a flag.
+  Cleanup moves to run CLOSE, which is what the checkout actually belongs to.
+  - **Watch citing epic-only code works by default.** Proved end to end rather than by construction:
+    `test/build-executor.test.ts` runs the real pipeline, then resolves a `[src: app:s1.txt:1]` against a
+    file the test first shows is committed on `epic/e1` and absent from the working tree.
+  - **Every close path takes them**, because a run does not only close one way: `tldrx next` closing the
+    last stage, `tldrx approve` signing the last gate, and `tldrx run cancel`. Without the last two the
+    change would have traded one leak for another — a cancelled run's checkouts used to be gone already,
+    because Build removed them on the way past.
+  - **`--keep-worktrees` keeps its meaning, one scope wider**: survive even the run close. It is
+    remembered on the run as `keep_worktrees:` (additive and optional; absent means clean up, which is
+    what every existing run.yml meant) because the flag is typed on the `tldrx next` that BUILDS and the
+    run is usually closed by a different command in a different process.
+  - Story worktrees are untouched: still removed the moment a story reaches `done` or `blocked`.
+
+- **The budget gate's three open policy questions are answered (#22, owner decision 2026-09-01).**
+  bb6204b wired the DATA — both economies, `attended_by`, `runSpend` — and deliberately changed no
+  verdict. These are the verdicts.
+  - **(a) An `attended_by: host` run is INFORMED, never DENIED, on metered dollars.** `tldrx next` on
+    such a run spawns nothing, so the estimate the gate was refusing against is spend that provably will
+    not happen. Both the PreToolUse hook and `tldrx next`'s own brake now say every number they would
+    have refused with, plus both economies, and allow. The event is `budget.warned`, not `budget.blocked`:
+    nothing was blocked, and recording a block that did not happen is the exact failure #22 was filed
+    about.
+  - **(b) A `host-tokens` ceiling is soft-enforced.** Under that economy the ceiling NUMBER is a
+    host-session token allowance, so accumulated declared `tokens:` against it is the one comparison in
+    the gate whose two sides share a unit. Crossing it WARNS and still allows. It stops only under the
+    explicit opt-in **`on_host_tokens_exceed: block`** in `budget.yml` — an enum beside `on_exceed`,
+    defaulting to `warn`, so every file written before the key existed keeps the behaviour it had. The
+    refusal never offers `tldrx budget raise`, which moves dollars and would send the operator at the
+    wrong number. (a) beats (b): the opt-in still never denies an attended run.
+  - **(c) `remainingWork` zeroes the developer share on an attended run**, mirroring `economy:
+    host-tokens`, because it is the same fact — the host session pays for those turns. Attendedness and
+    the phase economy were independent, so an attended run on a `metered-usd` phase still counted
+    developer turns against money this framework will never spend, on the brake and in `run estimate`
+    alike. Reviewer floors are untouched in both cases.
+
+### Added
+
+- **The documentation site speaks Spanish (`docs-site/es/`, phase 2).** All twelve English pages are
+  now mirrored under `/es/` at the same paths and translated into es-MX developer Spanish, and
+  `locales.es.themeConfig` carries the Spanish sidebar, nav, edit link and page chrome (outline,
+  prev/next, search modal, footer) rather than English chrome around Spanish prose. The placeholder
+  that phase 1 left at `/es/` is replaced by the translated landing page.
+  - **tldrx's own vocabulary stays in English where it is an identifier.** The stage names
+    (`What → How → Plan → Build → Watch`), `run`, `story`, `DoD`, `scope`, `handoff` and `workspace`
+    are the things you type or the files on disk, so they are not translated; each is glossed once in
+    Spanish where it first appears. `gate` is the one exception — it renders as **compuerta**, because
+    the phase-1 placeholder had already shipped that word to the live site.
+  - **Code blocks and command output are verbatim English** — a translated transcript would be a
+    transcript of a command nobody ran. The narration around them is translated.
+  - **Anchors were verified against the rendered HTML, not assumed**, because dead-link checking does
+    not see them. That mattered here: VitePress's slugifier strips accents but *keeps* `¿`, so
+    `## ¿Cómo lo detengo?` becomes `id="¿como-lo-detengo"`. The one heading that is linked to carries
+    an explicit `{#puedo-manejarlo-desde-claude-code}`. A sweep of the built site resolves all 5
+    anchor links across its 26 pages, 0 broken.
+  - **Release notes are deliberately not translated**: the page is generated from `CHANGELOG.md` at
+    build time, so a Spanish copy would drift. The `/es/` sidebar links the English page and says so.
+
+- **A public documentation site, written for people who have never seen tldrx (`docs-site/`, phase 1).**
+  A VitePress site deployed to GitHub Pages by `.github/workflows/docs.yml` on any push to `main` that
+  touches `docs-site/` or `CHANGELOG.md`. Twelve short English pages — a landing page, a Quickstart, one
+  page per concept (the five stages, files-as-state, gates, evidence, budgets), four guides and a
+  condensed CLI overview — plus a generated changelog and a Spanish placeholder. None of them is pasted
+  from `docs/`, which stays the agent-facing reference. Every command
+  and every block of output on the Quickstart was produced by running the real binary; nothing on the
+  site documents a flag that `--help` does not.
+  - **The changelog page is generated, never copied.** `docs-site/scripts/gen-changelog.ts` reads this
+    file at build time and emits one line per entry, so a release note reaches the website without
+    anybody maintaining a second copy of it. The generated page is gitignored for the same reason.
+  - VitePress dead-link checking is left ON and the build is green with it — proven by a probe, not
+    assumed: a deliberate link to a missing page failed the build with `1 dead link(s) found`. (Anchors
+    are NOT checked by it, so `#fragment` targets were verified against the rendered HTML by hand.)
+  - i18n is wired now, with the English content at the root and a Spanish placeholder under `/es/`, so
+    phase 2 is a matter of adding files rather than restructuring the site.
+  - `docs-site/` is excluded from the npm package (it is not in `files:`) and from `tsc --noEmit` (the
+    root tsconfig includes only `bin`, `src`, `test`) — measured: `npm pack --dry-run` still lists 52
+    files and none of them is under `docs-site/`.
+
+- **`tldrx learn` chapters 3-8 — the whole loop, played (#30, phase 2).** The tutorial now runs end to
+  end in about five seconds of real commands: **3** the gate (`approve --note`, and the record it writes
+  in `run.yml`), **4** one story built for real (How's `auto` gate closing itself over its seven
+  conditions, Plan's human gate, then a Build that cuts `epic/bulk-pricing`, spawns a developer in a
+  worktree, re-runs the story's `npm run test` DoD, commits, merges and spawns a reviewer), **5** a
+  genuinely red DoD and the three commands back from it (`story reopen`, `reject`, `budget raise`),
+  **6** an `agent` gate closed by `approve --as-agent` over a structured evidence note, **7**
+  `run attend host`, the refusal a bare `next` then gives, and the `next --prepare` / `next --commit`
+  pair actually run, **8** `cost --all`, `run estimate`, and the budget brake refusing a stage the
+  phase can no longer afford.
+  - Chapter 4's DoD is real, chapter 5's failure is real: the story's test script is `exit 0` until a
+    developer replaces it with a `node` test that then catches a wrong number — so the tutorial teaches
+    "a green DoD over an empty test proves nothing" by letting it happen rather than by saying it.
+  - Chapter 5 opens a second run, so `{run}` and `{runDir}` now expand in a step's `command` as well as
+    in a turn's writes, and mean **the newest run that is still open** — the same set `resolveRun` picks
+    from, so the placeholder and the CLI cannot disagree about which run a command means.
+  - Chapter 4's `prepare()` commits what `init` left untracked, because the Build executor refuses to
+    cut a branch from a dirty tree (measured: `?? .gitignore, ?? CLAUDE.md`, exit 2) — and the narration
+    teaches that refusal instead of hiding it.
+  - **Fixed: the toy repo now carries its own git identity.** Chapter 4's Build commits through the
+    framework's own executor, with whatever identity the machine has — so on a box with no global
+    `user.email` (a fresh laptop, a container, `ubuntu-latest`) `git commit` failed with `Author identity
+    unknown` and the chapter died three commands in. `makeSandbox` writes `user.email`, `user.name` and
+    `commit.gpgsign=false` into the sandbox repo's own config, on every open, so an older sandbox is
+    repaired too. The test pins it with `user.useConfigOnly` rather than an empty global config: git
+    guesses an identity from gecos and hostname and only fails where it cannot, so an empty config
+    passes on a laptop and fails in a container — which is exactly how this reached CI.
+
+- **`tldrx learn` — a playable sandbox tutorial that runs the REAL commands (#30, phase 1 of 3).**
+  A tutorial that runs the shipped binary can never drift from the shipped behaviour: every output the
+  learner reads is produced by the code, not written down by a doc author. `tldrx learn` scaffolds a
+  throwaway workspace (a four-file git repo with a `test` script that exits 0), then narrates, shows the
+  exact command, waits for Enter and RUNS it — `tldrx init`, `run new`, `next`, `answer` are the real
+  ones, against that sandbox.
+  - **Chapters 1 and 2 ship playable**: init (read the `workspace.yml` detection actually produced) and
+    the What stage (a question comes back, `tldrx answer` records it, it becomes `F001` in
+    `.tldrx/memory/facts.yml`). Chapters 3-8 from the issue are phase 2 and are DATA plus one
+    `assert()` — see the contract in `src/core/learn/Chapter.ts`.
+  - **It cannot spend money, by construction rather than by convention.** The sandbox writes its own
+    `claude` stand-in, names it in `TLDRX_CLAUDE_BIN` and puts it first on the child `PATH`, so neither
+    the spawn seam nor a bare `claude` on `PATH` can reach the real CLI. `test/learn.test.ts` proves it
+    the only way worth proving: it plants a booby-trapped `claude` that writes a marker file, plays both
+    chapters, and asserts the marker is absent AND the chapters completed — because a tutorial that
+    spawned nothing would pass a marker check for the wrong reason.
+  - **It cannot touch your work.** Everything is written under the sandbox directory (`~/.tldrx-learn`
+    by default, `--sandbox` to move it), and a sandbox that would sit inside a real tldrx workspace is
+    refused before a byte is written.
+  - **Files as state, like everything else**: `progress.json` in the sandbox is what makes a bare
+    `tldrx learn` resume, `--chapter <n>` jump (playing an unfinished prerequisite first), and `--reset`
+    start over. With no terminal on stdin the chapters play straight through rather than hanging at the
+    first prompt.
+  - The stand-in agent is scripted per chapter and is **fail-closed**: a prompt no turn matches exits 1
+    and names the turns it did have, so a hole in the tutorial is a loud failure rather than an
+    improvised answer. The `stream-json` writer moved from `test/fixtures/fakeStream.ts` to
+    `src/core/facilitator/fakeTranscript.ts` — beside `agentEvents.ts`, which reads that format — so the
+    tutorial's stand-in and the four test fakes cannot drift apart. The fixture re-exports it.
+
+- **`tldrx note <run> [--stage <id>] "text"` — an honest carrier for an operator annotation, at the
+  moment it happened (#46).** Measured on `260829-scoring-leaderboard` (2026-09-01): a host performed an
+  owner-delegated mechanical resync of eight story dod blocks, was asked to note it in the run log, and
+  could not — `events.jsonl` is append-only and tool-owned, so the only carriers were a FUTURE gate note
+  (late, and keyed to a decision the note is not about) or a `reject` (destructive). The session ended up
+  hanging the context off an unrelated `story.reopened`.
+  - **One event, and nothing else.** It appends a single `operator_note` line carrying actor, timestamp,
+    optional stage and the text. It does not go through `RunStore.save()`, which would rewrite `updated_at`
+    and re-derive every status for an annotation that changed no state: `test/operator-note.test.ts`
+    compares `run.yml` and `budget.yml` **byte for byte** across the call, because "safe to reach for
+    mid-run" is the whole of what makes the verb usable.
+  - **Every refusal writes nothing.** An unknown run (exit 3), a stage this run does not have (exit 2), an
+    empty note (exit 1) and a lone argument that turns out to name a run all refuse before the log is
+    opened. The last one is the trap worth naming: `tldrx note 260829-x` is a half-typed command, not a
+    note whose entire content is a run id, and recording it would be the one outcome nobody wanted.
+  - **Visible afterwards.** `tldrx run status` prints the last three (with `tldrx replay` named when there
+    are more), `--json` carries them as `operator_notes`, and `tldrx replay` narrates every one in place.
+
+- **`tldrx run gates set <stage>:<policy> --note "…"` — the signed upgrade path for a frozen
+  `gates_policy` (#14).** The policy is resolved at `run new` and frozen into `run.yml` by design, and that
+  default is not taken back here. What it left with no door at all, found on the 2026-08-30/31 unattended
+  pilots: a run created BEFORE the `agent` policy existed can never use `approve --as-agent`, and `run.yml`
+  is hand-edit-forbidden (spec §1) — so the only move was to abandon the run.
+  - **Built like `story reopen`, because it is the same kind of act:** a person overruling state the machine
+    is holding. A `--note` is required; ONE stage per invocation (a list would let a second change ride on
+    the first one's signature); the entry must be QUALIFIED, because under `--gates` a bare `plan` means
+    `human` and a signature must not rest on a default; and a no-op is refused rather than recorded.
+  - **One `gate.policy_changed` event** carries the phase, who signed it, the old and the new value, and the
+    note. A run with no `gates_policy:` at all gets the FULL map written — every stage explicitly, with the
+    one change applied — because a partial map would quietly claim its other stages had been decided too.
+  - It changes who may CLOSE a gate from then on. Gates already signed are untouched, and nothing re-reads
+    the policy of a closed one.
+
+- **`tldrx ship` — open a PR from the run's epic branch, with the handoff as the body (#15).** The loop
+  ended at "merge by hand": a finished epic sat on `epic/<slug>`, the document explaining it sat in
+  `<run>/<phase>/handoff.md`, and nothing carried either one to a PR.
+  - **It never pushes.** `core/build/git.ts` has no `git push` wrapper on purpose (spec §5), and this verb
+    keeps that rule rather than being the exception to it: a branch the remote has not seen is a refusal
+    that names the exact `git -C … push -u` command. Publishing a branch is a decision, and it stays the
+    operator's.
+  - **It never writes to the run** — no event, no gate, no cursor, no money — and it does not mirror
+    tickets: `tldrx tickets sync` already is that verb, holds the `process.yml` contract and appends
+    `ticket.synced`. A second, thinner mirror inside `ship` would give the workspace two answers to "is this
+    story mirrored", so `ship` names it as the next step instead.
+  - **Clean refusals, in a sentence:** no epic branch, no handoff, no `gh`, no remote, an unpushed branch,
+    several epic branches with no `--branch`, a branch the run did not cut. The body goes to `gh` as
+    `--body-file`, never as an argument, so a long handoff cannot overflow an argv limit.
+  - Both external binaries go through one narrow transport that takes a cwd — the same idea as
+    `adapters/transport.ts`, and the only way to ASSERT the argument shape of a command the suite must not
+    run. The unit tests drive a recording fake; the one end-to-end test puts a STUB `gh` first on PATH in a
+    throwaway workspace with a throwaway bare `origin`. The real `gh` is never invoked by a test.
+
+- **`TLDRX_CLAUDE_BIN` — point the sub-agent spawn at a different binary (#27, minimal slice).**
+  `spawnAgent` hardcoded `claude`, so a pinned install, a wrapper that adds a proxy or credentials,
+  and a stand-in in a sandbox all required patching source. The variable replaces the executable
+  NAME and nothing else — the argv is still Claude Code's, so what it points at has to speak
+  `-p --output-format stream-json --json-schema` — and blank or whitespace counts as unset. Honoured
+  everywhere the CLI is spawned: `spawnAgent`, the `--dry-run` command line (`describeSpawn`) and
+  `claude mcp list` (`McpProbe`). `tldrx doctor` deliberately still checks `claude --version`,
+  because `env.yml` declares that string. Documented under **Environment variables** in the CLI
+  reference. This is not the provider abstraction #27 asks for; #27 stays open for it.
+
+- **A drift guard on `templates/story.md` and `templates/epic.md` (#48).**
+  Both ship in the npm package, both state the Plan front-matter schema, and **nothing in `src/`
+  reads either one** — a second copy of a contract whose first copy is computed from `STORY_KEYS`
+  and `EPIC_KEYS`. Add a required key and `schemaContract.ts` stops compiling while the templates
+  say nothing; a human then opens one, writes a story the check refuses, and the framework looks
+  broken. They now go through `validateStoryFile` / `validateEpicFile` — the very checks the stage
+  gates on — with their key sets asserted equal to `STORY_KEYS` / `EPIC_KEYS` in order, and the
+  status enum each spells out in a comment asserted equal to `PLAN_STATUSES`. Proven to have teeth:
+  renaming one key and staling one enum comment turns three tests red. Whether the files should be
+  generated or deleted is a packaging decision and is left open on #48.
+
+- **The merge-wave sandbox is built under a hostile `init.defaultBranch` (#49).**
+  `test/merge-wave.test.ts` names every repo `main`, and CI run 33459567355 failed in the test's
+  own setup — `git push -q origin main` → `src refspec main does not match any` — on a runner whose
+  default branch is not `main`. `f1ffe56` had already fixed it (`-b main` on both inits,
+  `--branch main` on the clone), but nothing EXERCISED the fix: on a `main`-defaulting host,
+  removing the treatment changes nothing. The sandbox now pins `init.defaultBranch: trunk` for
+  every git command it builds itself with, the clone asserts it is on `main` rather than
+  discovering it five lines later, and two tests pin the mechanism — untreated reproduces the CI
+  error verbatim, treated does not.
+
+- **The Plan prompt now STATES the schema the `plan` check enforces, generated from the check itself (#35, #38).**
+  `stages/plan/stage.md` named the output filenames — `stories/<id>.md`, `epics/<epic>.md`, `waves.yml` — and
+  said nothing about their shape, so a fresh agent learned it by having a paid attempt refused. Measured twice
+  in two days: on `260831-hardening-d1` the plan sub-agent followed the rendered bundle faithfully and wrote
+  seven stories as plain markdown (`no YAML front matter — the file must open with ---`), and on
+  `260829-scoring-leaderboard` it wrote a 1,009-character acceptance item against a `MAX_ITEM_CHARS = 512`
+  cap that appeared in no file it could read. Both attempts were consumed, correctly and uselessly.
+  - **Generated, not copied.** `src/core/plan/schemaContract.ts` renders a `## Output schemas` section from
+    `STORY_KEYS`, `EPIC_KEYS`, `PLAN_STATUSES` and the six `MAX_*` constants — the same definitions
+    `validateStory`, `validateEpic` and `validateWaves` read. `Record<StoryKey, Field>` is load-bearing:
+    add a key to a schema and the file stops compiling until the new key has a value and a rule. The worked
+    examples it ships are run through `validatePlan` itself in the tests, so the contract the prompt states
+    and the contract the check enforces are provably the same one. This repo already had the other kind:
+    `templates/story.md` carries the schema correctly and `grep -rn 'story\.md' src/` finds nothing that
+    reads it.
+  - **Only where the check runs.** `applyCheckContracts` splices it under an H2 the framework owns, for a
+    stage that declares `checks: [plan]` AND writes `waves.yml` — the same predicate `checkPlan` skips on,
+    now shared (`writesPlanArtefacts`). A What or How prompt is byte-identical to before. It goes into
+    `stage.md` rather than after the inputs because `prompt.ts` orders the document most-stable-first for
+    the prompt cache, and a section computed from constants is exactly as stable as the stage body.
+  - **~4.2 KB against a $4 stage.** The alternative it replaces is a refused attempt per fresh workspace.
+
+- **`tldrx plan sync-dod` — the mechanical repair for dod blocks an edited `workspace.yml` orphaned (#42).**
+  A story's ```dod block may only name commands `workspace.yml` declares, byte for byte, and that rule is not
+  relaxed by a byte here — it is what stops a data file from running an arbitrary command as you. What it
+  lacked was an inverse. Measured live 2026-08-31 on `260829-scoring-leaderboard`: fixing `workspace.yml`
+  (a filtered `test:`, `lint:` deleted) instantly invalidated the dod blocks of **8 approved stories**, and
+  the only recoveries were hand-editing agent-approved artefacts or re-running the whole Plan stage — a paid
+  turn to change two lines in eight files, churning thirteen correct stories on the way.
+  - **Evidence, not similarity.** The ancestry comes from git's history of `.tldrx/workspace.yml`: a line a
+    previous version declared under a role the current file still has becomes that role's current command; a
+    line whose role is gone is dropped; a line the current file already declares is left alone; and a line no
+    version ever declared — or one two roles once shared and now disagree on — is **flagged, its story left
+    byte-identical**, and the command exits `2`. Guessing at a rename by string shape is the one thing this
+    must not do. In a workspace with no git history there are no ancestors, so everything non-current is
+    flagged.
+  - **It touches nothing else.** Front matter, prose, blank lines and the fences come back byte-identical;
+    only the command lines inside the fence move. The previous version is kept at `<story>.md.bak`
+    (`writeAtomic`), the summary is a per-story diff, `--dry-run` prints it and writes nothing, and the result
+    is validated by the same plan check the drift came from. Stories that CAN be synced still are — one
+    undecidable line is not a veto on the other seven files.
+  - **The drift message now names the remedy.** "`<cmd>` is not one of .tldrx/workspace.yml's commands — a
+    story may not invent one" gained a second sentence pointing at `tldrx plan sync-dod`. Only for a STORY:
+    a stage's `cmd:` is a line a human wrote, and `sync-dod` does not touch stage files.
+
+- **`tldrx answer <Qn> "…" --supersede` — the verb for reversing a decision already on record.**
+  Found live 2026-08-31: an owner reversed an answered decision after the risk behind it was
+  refuted, and `tldrx answer` refused ("Q1 is not an open question") because an answer is
+  recorded once. `superseded_by` had been in the §2.5 schema since the first draft with **no
+  command that wrote it**, so the only route was a hand edit of `facts.yml` — and a hand edit
+  that left `superseded_by: null` left the reversed decision inside `FactsStore.active`, which
+  every stage reads as never-re-ask truth. The next run would have reinstated the call the owner
+  had just taken back.
+  - **The verb.** Valid only on an ANSWERED question (on an open one it exits `1` and says to
+    answer it normally; without the flag an answered one still exits `3`, now naming
+    `--supersede`). It appends a new fact carrying the whole new answer with the same `area` and
+    `repos` and ordinary provenance, sets the old fact's `superseded_by` and the new one's
+    `supersedes` — both halves, through `FactsStore.supersede`, under the workspace lock, so the
+    reciprocity rule cannot be broken — and never edits a byte of the old fact's text. Reversing
+    twice supersedes the SECOND fact: the chain is walked to its head, so it stays single-link.
+  - **The questions block is appended to, not rewritten.** The original `[Answer]:` line and its
+    footer stand; a superseding answer line and a `reanswered_by | reanswered_at | fact |
+    supersedes` footer go under them. `status:` stays `answered`, because it is.
+  - **Every reader that feeds a decision now skips a superseded fact.** This was half the work
+    and none of it was new code: `superseded_by` had a writer for the first time, and six readers
+    had been filtering on retirement alone. `isLive` (`core/facts/Fact.ts`) is now the one
+    predicate behind `FactsStore.active` (no-re-ask, `tldrx run new --from` de-duplication),
+    `findDuplicate`, `renderFacts` (the `{{facts}}` section of **every** prepared prompt),
+    `renderWatchFacts`, `runFacts` (the implicit plan's "this run's answers") and `relevantFacts`
+    (the training miner). One test had pinned the old behaviour in words — "a
+    superseded-but-not-retired row stays visible" — and that was the bug, not the rule.
+  - **History readers still show it, labelled.** `tldrx replay` renders the new `fact.superseded`
+    event as its own line — the one moment the workspace's durable memory changes its mind was
+    the one moment replay could not narrate — and `tldrx retro` lists the old fact with
+    `(superseded by F<n>)` beside it.
+
+### Fixed
+
+- **`tldrx learn` — the cold-player QA round (#30).** A first-time player played all eight chapters
+  and returned SHIP-with-fixlist. Everything they found is fixed or recorded:
+  - **Chapter 8 no longer lies about the brake.** It said "the phase has already spent its Watch
+    money" while the tool printed `$1.89 left … estimate is $2.00`. The real mechanism is that a
+    re-run is priced at the stage's WHOLE declared `budget_usd`, never at what a second attempt might
+    add — so a stage that has spent anything can no longer afford itself. The chapter now says that,
+    quotes both figures, and `assert()`s them against the `budget.blocked` event so the numbers cannot
+    drift away from the sentence.
+  - **Chapter 1 no longer promises something chapter 2 does not deliver.** `--no-interview` skips
+    *init's* setup interview, which no chapter covers; the forward reference is gone and the debrief
+    now sends the learner to `.tldrx/init-handoff.md`, where measured/inferred/assumed and
+    `[src: …]` / `absent:` actually live.
+  - **`tldrx learn --chapter <n>` refuses a chapter that is already played**, up front and by name,
+    instead of narrating it and then dying mid-chapter on `run new: … already exists` (exit 1,
+    measured). The refusal names `--reset` and the chapter a bare `learn` would resume at.
+  - **The tutorial has a door out.** The ending now names the first four commands to type on a real
+    repo — `tldrx init` (with the warning that it runs an interview by default), `run new --scope
+    hotfix`, `next`, and `tldrx ship`.
+  - **Chapter 7 RUNS `next --prepare` and `next --commit`** against the feature run's Watch stage
+    instead of describing them in a debrief. Chapters 6 and 7 swapped for it: the attended chapter
+    addresses the feature run through `{run}`, so the hotfix run has to be signed off first.
+  - **Every non-zero exit code is printed** (`→ exit 4`), so the code chapter 2 teaches is a thing
+    the learner reads rather than a thing they are told. Chapter 5 now also demonstrates the exit-2
+    refusal a bare `next` gives with two runs open, and names the two run-id spellings.
+  - Jargon defined at first use — expert, level 0, the `claim-sources` / `no-reask` / `budget-gate`
+    bracket, `boundary`, `[src: …]`, `absent:`, economy, §2.11 — and the `expert … has no evidence`
+    nudge explained once instead of repeating unexplained nine times.
+  - Known and NOT fixed: a step's stderr (where the agent stream lives) is buffered and printed after
+    its stdout, so a summary can appear before the stream that produced it. Interleaving needs an
+    `onStderrLine` on the runtime seam and in both implementations; documented in `engine.ts`.
+
+- **`tldrx cost` no longer claims "two economies" over one (#56).** The `(no total: two economies, no
+  exchange rate)` footnote was unconditional, so a run whose every attempt was metered in dollars was
+  told no total could be printed. It is printed only when both economies are actually present.
+
+- **The README's "Not on npm yet" warning was false and told readers not to run the install line
+  directly underneath it.** The package IS published: `npm view tldr-experts version` → `0.3.1`,
+  exit 0. The warning is removed rather than re-dated — the npm badge at the top of the README
+  already shows the live version, so nothing in its place can go stale the same way. The
+  `npm i -g tldr-experts` line it was contradicting is unchanged.
+
+- **The site's own home page linked an anchor that does not exist.** The hero's "Try it offline,
+  free" button pointed at `/quickstart#try-the-whole-thing-first-for-free`, but the heading renders
+  as `id="first-try-it-for-free"` — verified against the LIVE page, not just a local build. Dead-link
+  checking never saw it because VitePress does not check fragments. Repointed, and a sweep of the
+  built site now resolves every anchor link it emits.
+
+- **`tickets sync`, `tickets status` and `budget show` took a run id as a positional that neither
+  their `usage` nor their `--help` declared (#53).** Measured at `7ac298c`:
+  `tldrx tickets status zzz-positional-probe` and `tldrx budget show zzz-positional-probe` both reach the
+  run resolver and exit `3` with `no run 'zzz-positional-probe'`, so both forms have always been
+  supported. The mechanism is the same in each: the subcommand word is consumed by the dispatcher
+  (`tickets.ts:53`, `budget.ts:32`) before `stringFlag(args, "run") ?? args.positionals[0]` runs
+  (`tickets.ts:246`, `budget.ts:48`), so `positionals[0]` is a run id by then.
+  - **The capability is DECLARED, not removed.** `usage` now reads `tldrx tickets sync [<run>] …`,
+    `tldrx tickets status [<run>] …` and `tldrx budget show [<run>] …`, and both help entries gain the
+    `[<run>]` arg every other run-scoped command already carries. Nothing about what the CLI accepts
+    changed.
+  - **This is the axis #51's guard cannot see.** That guard compares the registry to the usage; here the
+    registry itself was narrower than the code, and where both are silent both are green. The new check
+    in `test/cli.test.ts` is a hand-written list — nothing derives a positional from source — but its
+    BEHAVIOURAL half spawns the real CLI against a throwaway workspace, so it also goes red if the
+    capability is ever removed, which is the direction a tidy-up of the arg parsing would break it in.
+
+- **`tldrx run gates set` was documented nowhere in `docs/` (#54).** `grep -rn "gates set" docs/` returned
+  nothing at `7ac298c`, so the CLI reference — the page a reader lands on from the README — described
+  seven of `tldrx run`'s eight subcommands. It matters more than an ordinary docs gap because `gates set`
+  is the ONLY sanctioned way to move a `gates_policy` that `run new` froze, and the situation it exists for
+  (a run opened before the `agent` policy existed, which can otherwise never use `approve --as-agent`) is
+  one an operator hits mid-run and searches the docs for. What they found was "abandon the run".
+  - Documented in all three places the question gets asked from: the `tldrx run` usage block and a new
+    prose entry in `docs/guide/08-cli-reference.md`; a **Moving a frozen policy** section in
+    `docs/guide/03-runs-and-gates.md`, right under the paragraph that explains the freeze, with the
+    `--gate-agent` disclaimer further down now linking to it; and `docs/spec.md`, both in §2.2's
+    `gates_policy` row and as its own §CLI row (exits `0,1,2,3`, each measured).
+  - Every copy carries the two facts a usage line cannot: **`--note` is mandatory**, and the change
+    appends one **`gate.policy_changed`** event with the actor, the moment, the note and the old→new
+    value — the whole audit trail for a mutation nobody would otherwise go looking for.
+  - Guarded: `test/cli.test.ts` now asserts the CLI reference names every subcommand in
+    `runCommand.subcommands`, plus `gate.policy_changed` by name. Scoped to `run` on purpose —
+    `plan sync-dod` has no section on that page at all (a separate gap, unfiled), and `hook`'s seven
+    scripts are deliberately documented as one `<script>` slot.
+
+- **The epic file duplicated every story's status, and nothing ever updated the copy (#50).**
+  Measured on `260829-scoring-leaderboard` (2026-09-01): `03-plan/epics/E1.md` listed S1, S2 and S3 as
+  `todo` in its `## Stories` table while `03-plan/stories/S1.md` said `done` (merged at `0a50660`,
+  `task.done` in `events.jsonl` at 23:46:11Z) and S2 said `in_progress`. Nothing had lied — nothing had
+  written, either. Fixed by **removing the copy**, not by adding a second writer, and the repo already
+  drew that line for the one field it does maintain: the epic's front-matter `status:` is DERIVED and
+  written by `BuildExecutor.updateEpicStatus` (`build.ts:2552`) from the story files. A copy with a
+  writer is a cache; a copy without one is a lie waiting to be read.
+  - **Nothing parsed the table.** `validateEpicFile` is front matter only ("the front matter is the whole
+    schema", `schemas/epic.ts:69`), `adapters/body.ts:50` mirrors an epic to a ticket as a bare list of
+    ids with no status, and the dashboard reads the front matter. A writer would have been maintaining a
+    document with no reader.
+  - **Both copies of the shape are fixed, not just the visible one.** `templates/epic.md` now points at
+    `03-plan/stories/<id>.md` instead of tabulating it, and the GENERATED contract the Plan sub-agent
+    reads (`schemaContract.ts`, spliced into the stage prompt) now says "Do NOT restate a story's status,
+    repo or `depends_on`" — without that, the next Plan agent invents the table again, which is how it
+    got there.
+  - `test/plan-schema-contract.test.ts` runs the issue's acceptance: build a plan from the shipped
+    templates, flip S1 to `done` through `updateStoryFront` (the writer the Build executor uses), then
+    grep the epic. Any `S<n>` + status word on one line of the epic body is a claim, and the claim set
+    must be empty.
+
+- **Seven `usage` strings were narrower than the same command's `--help` (#51, after #25).**
+  `usage` is what a BAD invocation prints — `run.ts:85`, `questions.ts:38`, `tickets.ts:60`, `gate.ts:44`
+  and three more write `<cmd>.usage` to stderr — so it is the string an operator reads at the exact
+  moment they got the invocation wrong, and it was hiding flags the code accepts. Widened: `run attend`,
+  `run status`, `run estimate`, `run auto`, `run unlock` and `run cancel` now show `[--run <id>]`
+  (all six read `args.positionals[0] ?? stringFlag(args, "run")`); `tldrx next` shows it too
+  (`next.ts:48`); and `tldrx questions lint` shows the `[<run>]` positional it has always taken
+  (`questions.ts:49`, and `docs/guide/08-cli-reference.md` had been documenting it for longer than the
+  CLI admitted it).
+  - **The guard is subcommand-aware, and that is not gold-plating.** A plain
+    `usage.includes("--run")` calls `run` CLEAN, because the new `run gates set` line names `--run` —
+    measured, the naive check saw **one of run's seven gaps**. `test/cli.test.ts` scopes a flag that
+    declares a `sub:` to that subcommand's block of the usage, and falls back to the whole string for a
+    `sub:` that is a MODE rather than a word in argv (`dashboard --out {sub: "static"}`, which the first
+    draft reported as a gap it is not).
+  - **Three of the seven were spelling, not gaps, and are allowlisted with the reason:** `seed`'s
+    `<Qid> "<text>"`, `watch`'s `check <feature>` and `hook`'s enumerated script names all say the same
+    thing more specifically than the registry's general name. A fourth, `run`'s
+    `<stage>:<human|auto|agent>`, is the same case. The allowlist is itself checked: every entry must
+    still name a declared flag or arg, so a rename turns an exemption red instead of silent.
+  - **`tickets --dry-run` was left OUT on purpose.** `tickets sync` previews by default and `--apply` is
+    the write; advertising `--dry-run` would imply the opposite, and `test/money-safety.test.ts:319`
+    asserts its absence. Recorded in the allowlist as a decision rather than papered over as a gap.
+
+- **A literal ESC byte in `McpProbe.ts`'s ANSI regex (#52).** `src/core/doctor/McpProbe.ts:11` wrote a raw
+  `0x1b` where `\x1b` was meant, so the source read `/<ESC>\[[0-9;]*m/g` and a reader — in a diff, in a
+  review, in a terminal, in most editors — saw `/\[[0-9;]*m/g`, a different and wrong-looking regex that
+  someone tidying is one keystroke from breaking `tldrx doctor --mcp` with. **Not the #47 hazard**, and
+  worth saying: ESC does not trip the binary-file heuristic, and the file was always visible to a grep.
+  Behaviour is byte-identical, and `.source` is the wrong instrument for proving that (it returns the
+  literal as written, so the two spellings differ there while compiling to the same matcher) — so
+  `test/doctor.test.ts` compares the shipped `stripAnsi` against a reference rebuilt from the old
+  literal-ESC form over a nine-line corpus, 4 of which change. `test/source-hygiene.test.ts` now flags a
+  raw ESC as well as a NUL: measured across all 479 `.ts` files under `src/`, `test/`, `bin/` and
+  `scripts/`, `McpProbe.ts` held the only one, so the check has no false positives to trade against.
+
+- **A stray NUL byte made two source files invisible to every grep-based sweep (#47).**
+  `test/cli.test.ts` carried one literal `0x00`, so `file(1)` called it `data` and `grep -I` —
+  ripgrep and ugrep too — dropped it SILENTLY, exit 0, no message. Measured on `origin/main`:
+  `grep -lI -E 'node:child_process|Bun\.spawn' test/*.ts` returned **36 files with `cli.test.ts`
+  absent**, though it calls `Bun.spawn` on line 32. That is how it missed #43's load-aware timeout
+  and then timed out at 5004 ms on the very merge that was fixing timeouts. Writing the guard found
+  a **second** one nobody had reported — `src/core/text/srcToken.ts:711`, a NUL used as a cache-key
+  separator, which hid that file from every `src/` sweep (367 of 368 `.ts` files visible). Both are
+  now the two-character escape `\0`: identical at runtime, ordinary text on disk. Post-fix the same
+  sweep finds `cli.test.ts` and all 368 `src/` files. `test/source-hygiene.test.ts` walks `src/`,
+  `test/`, `bin/` and `scripts/` and fails on any NUL, with the offender named at `path:line`.
+
+- **The five wave-5 docs-pass nits, each a sentence nothing was checking (#25).**
+  - **`boundary.ts` promised an exclusion is "never silent" and dropped state paths without a
+    word** — `BoundarySurface.excluded` was populated and read by nothing. Every verdict that has a
+    surface now names what was excluded, green and red alike, including the case where the
+    exclusion was ALL there was and the run therefore reported "declares no surface".
+  - **The precondition refusal asserted "the stage is still `ready`" without looking.** It reports
+    the status `run.yml` actually holds. Bigger than filed: on a FRESH run the stage at the cursor
+    is `pending`, not `ready`, so the old sentence was wrong in the ordinary case as well as on the
+    retry of a `failed` stage.
+  - **The agent-gate fallthrough printed its label twice** — `boundary: boundary=…`, because every
+    condition detail was prefixed with its own id including the two that have a trigger of their
+    own. Only the generic `condition` trigger keeps the prefix; alone it names nothing.
+  - **`dispatchNotes.ts` documented `.agent/04-build/build/S5/…`**, one phase segment more than
+    `dispatchNotesPath` builds. The example is now asserted equal to the path the code produces.
+  - **Two usage strings were narrower than their own `--help`**: `tldrx gate template` omitted the
+    positional `[<run>]` it accepts, and `run new` spelled `--gates <a,b|all|none>` where the help
+    says `<a,b|a:agent|all|none>`.
+
+- **`tldrx next --dry-run` spawns nothing. It used to cost $0.42 a go (#17).** The flag ran the
+  stage for real — one `claude -p`, one `agent.spawned`, one `agent.result`, the cost on the
+  ledger — and only reverted the non-handoff FILES afterwards. Measured on the 2026-08-30 pilot;
+  `tldrx next --help` had said "Spawns nothing and writes nothing" the whole time, so this is the
+  code catching up to the promise rather than the promise being watered down to the code.
+  - **What it does now.** It assembles the prompt, prices it, and stops: the expert bundle, the
+    context ledger, the prompt size, the declared outputs, and the **exact `claude -p` argv** it
+    would have run (with the `--json-schema` blob elided as `<envelope-schema>`, rendered from
+    `buildClaudeArgs` itself so the printed command cannot drift from the real one). Then the two
+    commands that would actually dispatch it. Exit `0`.
+  - **Nothing is written either.** No prompt bundle and no `pending.json`, so a dry run cannot
+    leave a `--commit` looking at a turn that never happened; the stage keeps its status and the
+    ledger keeps its zero. `dry_run_allowed: false` still refuses (Build sets it: a stage that cuts
+    branches and fans out per-story sub-agents has no ONE dispatch to describe).
+  - **On an attended run it is still refused at exit `4`** — but for the right reason now. It costs
+    nothing; it describes a dispatch the framework never makes there, and `--prepare` writes the
+    bundle the host is going to carry. The message said "it spawns a real sub-agent" and no longer
+    lies.
+
+- **A precondition gets its own clock, not the stage's 900–1800 s (#20).** `preconditions:`
+  inherited `timeout_s`, so one hung command — `docker info` against a dead daemon is the measured
+  case — could hold a run for half an hour: exactly the waste the feature exists to prevent, taken
+  by the guard instead of by the attempt. Each precondition now gets **60 s** by default
+  (`PRECONDITION_TIMEOUT_S`), overridable per entry with `timeout_s: <n>`, refused at load if that
+  is not a number `> 0`. A timeout is a red precondition like any other — exit `2`, nothing
+  written, nothing spawned, the stage where it was — and its message names the precondition, its
+  own timeout and the knob that changes it, rather than the stage's. `CommandRun` gained
+  `timedOut` so a timeout can be told from a refusal or a wrong exit code without reading prose.
+
+- **The budget gate can see host-token spend and attendedness (#22).** The tolerant reader the
+  `budget-gate` hook and the status line share (`hooks/lib/runFile.ts`) skipped `tasks[]` and
+  `attended_by:` entirely. So a run whose turns a host session paid for reported `$0.00` metered
+  and nothing else, and `runSnapshot`'s tolerant path hard-coded `attendedByHost: false` with a
+  comment admitting it meant "cannot see". `RunView` now carries `attended_by` and each task's
+  `cost_usd` / `metered` / `tokens`; `runSpend` derives the metered dollars, the declared host
+  tokens and the uncosted turns; `renderRunEconomies` renders the one line that says a dollar
+  figure is a lower bound. The gate appends it to a `host-tokens` phase's stderr note and to a
+  refusal, and `budget.blocked` records `economy`, `attended_by`, `metered_usd`, `host_tokens` and
+  `unmetered_tasks`. **No verdict changed**: a dollar ceiling still governs dollars, the two
+  currencies are still never converted, and a plain metered run's refusal is byte-identical.
+
+- **The `max_reads` flake was a real race, not a slow test (#24).** A chunk boundary is not a line
+  boundary: `LineSplitter` hands every complete line in one chunk to the read counter
+  synchronously, so when the OS coalesced the sub-agent's writes — which is what a loaded CI box
+  does — reads 4..20 were counted in the same tick as read 3, long before the `SIGKILL` just
+  ordered could land. `agent.result.payload.reads` was therefore a function of scheduling, and the
+  assertion pinning it to the cap cost two retries in one night. The counter now stops the moment
+  the cap fires, so what is recorded is the number of reads the cap ALLOWED. Pinned by a fixture
+  that makes the coalescing deterministic (`FAKE_CLAUDE_READS_BURST=1` — every read pair in one
+  write): pre-fix that reported 20 reads against a cap of 3.
+
+- **One over-cap list item no longer cascades into false "S<id> has no file" errors (#37).**
+  `validatePlan` resolves cross-file references out of the set of stories that PARSED, so a story file that
+  failed its own validation was indistinguishable from one that was never written. Measured on the
+  `260829-scoring-leaderboard` session: `acceptance[3]` in `S8.md` was 1,009 characters against the 512 cap,
+  and the check reported three errors of which one was real — the other two said `S8 has no file in stories/`
+  about a file that was 5,794 bytes on disk. The operator only avoided a wasted pass by re-deriving the cause
+  from the validator source; an agent reading that message goes hunting for a missing file or rewrites
+  `waves.yml`.
+  - A reference to a story or epic whose FILE EXISTS is never reported as missing. It now reads
+    `S8 is unresolved because stories/S8.md failed validation — that file exists; fix the errors reported
+    against it and this one goes with them`, carries `cascade: true` on the `PlanIssue`, and covers the
+    id-mismatch case (`stories/S8.md declares id \`S9\``) as well as the invalid-file case. The epic side —
+    a story pointing at an epic whose own file did not validate — had the identical bug and the identical fix.
+  - `describePlanIssues` orders root violations ahead of cascades. Its window is three issues wide, so one
+    real defect cascading into four references could otherwise spend the whole window on consequences and
+    never name the cause.
+  - A story that is genuinely absent still reports `has no file`, with no cascade flag.
+
+- **A refused list value now names the cap it broke, at the cap's current value (#38).**
+  The constants were interpolated already but the sentence was not self-describing: `513 characters exceeds
+  the 512 cap` did not say the cap is per-item or that splitting the item is the fix. Now
+  `513 characters exceeds the 512-character cap on one list item — split it into several items` and
+  `65 items exceeds the 64-item cap`, both still derived from `MAX_ITEM_CHARS` / the list's own `max`, and
+  the same constants are what the Plan prompt states up front.
+
+- **The merge itself is now serialised, and a gate can no longer describe a tree it is not pushing
+  (#44).** `scripts/merge-wave.sh` merges, gates and pushes in ONE shared checkout and took no lock.
+  Measured on the pre-fix script with two concurrent invocations against a real sandbox repo: run A
+  gated `7afcc0e` (its own merge) at `typecheck` and `fadc923` (the OTHER run's merge, landed
+  mid-gate) at `build`, then printed `OK fadc923 … pushed`. Both runs reported the same sha and
+  both exited 0 — a green report over a tree neither had finished gating. With a red change in the
+  other branch the same interleaving hands agent A a `FAIL build=1` for code it never wrote, which
+  is what was actually observed live 2026-08-31 (`2184` tests counted where the branch had `2181`).
+  - **A lock, held from before the dirty-tree check through the push.** `mkdir` on
+    `.git/merge-wave.lock` — atomic on macOS and Linux, where `flock(1)` is not on stock macOS, and
+    in `git rev-parse --git-common-dir` so the lock can never be dirt in the tree it guards — and
+    not `$R/.git`, which in a linked worktree is a FILE that `mkdir` can never turn into a lock. A second invocation WAITS,
+    saying so on stderr (`merge-wave: waiting for another merge in this checkout (owner: …)`) so the
+    single summary line on stdout stays a single line. Waiting is bounded (`MW_LOCK_WAIT_S`, default
+    3600 → exit `6`), and a lock whose owner is a dead pid on this host, or older than
+    `MW_LOCK_STALE_S`, is broken open — after re-reading the owner line, so two waiters cannot tear
+    down a lock a third has just taken. An interrupted run hands the lock back on its way out: an
+    untrapped signal kills bash WITHOUT running its `EXIT` trap, so `INT` and `TERM` are trapped too. No
+    path through the wait loop is free of the budget, including the break-open one: a lock that
+    cannot be created or removed now fails in under a second instead of spinning forever.
+  - **And an assertion that does not depend on the lock.** Between the last gate and the push, HEAD
+    must still be the commit the gates ran against; if it moved, nothing is pushed and the script
+    exits `5` saying which sha it gated and which one is there now. The lock prevents the race; this
+    makes pushing an ungated HEAD impossible even for someone who bypasses the lock. The pre-fix
+    script, given the same mid-gate commit, pushed it and reported `OK`.
+  - **It pushes the commit it gated, not the `main` ref.** `git push origin main` publishes
+    `refs/heads/main` whatever HEAD is — and a red gate leaves `main` sitting on an ungated merge
+    commit by design, so the next run from a detached or repaired HEAD would have published THAT.
+    The push is `HEAD:main` now, and a pre-flight refuses (exit `7`) when the gated commit is not a
+    fast-forward of `origin/main` rather than letting the server's rejection be the first news.
+  - Gate logs moved from the fixed `/tmp/mw-*.log` to a per-invocation `${TMPDIR}/mw-<pid>/`, and the
+    FAIL lines name the directory. Two runs in two clones on one box shared those files.
+
+- **The test suite no longer goes red because the machine was busy (#43).** On an untouched
+  `origin/main`, `bun test` reported `2155 pass · 5 fail` while the same two files alone reported
+  `91 pass · 0 fail`: four tests that spawn a REAL `git` expiring on bun's 5000 ms default, and one
+  50 ms performance budget measured at 66.4 ms, with three `tldrx` runs and two other agents sharing
+  the box. Because `merge-wave.sh` refuses to push on any test failure, that red is indistinguishable
+  from a regression at the exact moment a merge is decided, and the natural response — re-run until
+  green — is how a real regression eventually gets pushed.
+  - **The clock moved; no assertion did.** `test/fixtures/machineLoad.ts` measures the machine
+    (1-minute run-queue per core, floored at 1 and capped at 8) and hands out budgets from it. All
+    **42** test files that spawn a real process — `git`, `bun`, the CLI — now open with
+    `setDefaultTimeout(spawnTestTimeout())`: 30 s idle, scaled by load, still a hang detector. How
+    long a process takes to start is a property of the machine, not of the code, so a fixed budget on
+    such a test measures the box. A test enumerates those files and fails if a new one skips the
+    budget. Serialising the suite would not have helped: `bun test` already runs files sequentially
+    in one process (verified — a `setDefaultTimeout` in one file does not reach the next, and a 5.5 s
+    test in that next file still expired at 5000 ms). The contention is other processes on the box,
+    which only a load-aware budget can see.
+  - **The first version of that list was a `grep -l`, and it lied.** It returned 14 files and silently
+    omitted `cli.test.ts`, whose "every command's help lists an exit table" then timed out at 5004 ms
+    on the merge that was fixing timeouts. Cause: one stray NUL byte at `test/cli.test.ts:366` makes
+    the file `data` to `file(1)`, and grep drops binary files under `-I` without a word. The list is
+    built by READING every file now, and `cli.test.ts` is asserted to be in it. Filed as #47.
+  - **The one real performance budget keeps its teeth.** `handoff` on 256 KB is now the floor of
+    three runs against `perfBudgetMs(50)`, which on an idle machine is 50 — the identical assertion.
+    A stall inflates some runs and never the floor, and a function that genuinely takes 120 ms still
+    fails, which is itself a test.
+
+- **`npm pack` output no longer refuses the next agent's merge (#45).** `tldr-experts-<version>.tgz`
+  was not ignored, and the dirty-tree guard refuses on ANY porcelain line, untracked included — so a
+  pack artifact left by a release check blocked the merge of whoever came next, someone who did not
+  create the file and could not know whether deleting it was safe. `*.tgz` is ignored now. The guard
+  is deliberately unchanged: an untracked file is still dirt, and a test holds it to that.
+
+- **`claim-sources` reports every problem it found, over every declared `.md` output — and a
+  `file` src resolves against this run's epic worktree.** Four issues, one code path (#33, #34,
+  #23, #16), all four measured on the 2026-08-30/31 unattended pilot runs.
+  - **It reported ONE problem** (#33). `checkClaimSources` returned on the first file, the first
+    category and `unresolved[0]`, so a 226-bullet cap breach sat invisible behind a single bad
+    file path: fixing the visible one and re-running would have bought the next one at the price
+    of a full paid pass. Every file and every category is reported now, as a per-file summary
+    (`<file>: 3 unsourced bullet(s) on line(s) …; 2 unresolvable source(s) — …`), with file-level
+    problems such as the cap breach listed FIRST so 200 line numbers cannot bury them. Up to six
+    of a category are named and the rest become `(+N more)` — the same convention
+    `describeKnowledgeIssues` uses, and necessary because a check's `detail` is rendered inside
+    one-line summaries (`autoGate`, `next`).
+  - **It looked at ONE file** (#34). The filter was `endsWith("handoff.md")`, so the identical
+    violation refused the stage when it was written in `handoff.md` and passed in silence when it
+    was written in `design.md`, `contracts.md` or `scope.md` beside it — the pilot's pass-3
+    violation was caught only because it happened to be in the handoff. Every declared `.md`
+    output is read now, by both the gate check and the write-time hook: the four-section rule for
+    the files that ARE handoffs, and `validateCitations` for the ones that are not. That second
+    rule is deliberately narrower — a bullet with no citation is prose, but a `[src: …]` that WAS
+    written must parse, must resolve, and must obey `$ … → exit n` belonging only to an
+    `Evidence ledger`. A declared non-handoff output that was never written is still not a
+    failure; that is the `--commit` gap check's job.
+  - **The execution-claim validator reads the verb** (#23). `\bexit \d` missed "exits 0", which
+    is how a trainer writing normal English says it, so the claim slipped through the grammar the
+    rule exists to enforce while "exit 0" three words away was refused. Conjugation, an optional
+    "with", and the `code`/`status` spellings all match now; the digit is still required, so "the
+    exit path is documented" and "the exchange refuses an empty code" stay prose.
+  - **A `file` src resolves against this run's epic worktree** (#16). The Build phase commits
+    onto an epic branch and deliberately does not merge it, so a Watch-stage handoff ABOUT that
+    work had every `repo:src/…` citation refused for naming code the working tree does not have
+    yet — the stage's own evidence was rejected for being true. `.tldrx/worktrees/<repo>/_epic-<run>-<epic>`
+    is now a resolution base, tried before the working tree, for both the hook and the gate; the
+    path convention has one home (`core/paths.ts`) that the Build executor writes and the §2.8
+    resolver reads. Resolution also no longer stops at the first base where the file EXISTS but
+    is too short — a file truncated on the epic branch would otherwise deny a claim about the
+    line it still has on `main`. **Still open** (commented on #16): the epic worktree is removed
+    at the end of Build unless `--keep-worktrees`, so the default Watch stage has no tree on disk
+    to resolve against. Closing that means reading blobs out of the epic branch inside a hook
+    whose budget is 50 ms, or keeping epic worktrees for the life of the run — a design call, not
+    a mechanical one.
+
+- **The Build DoD is a DELTA gate again: the base tree is checked before any story is charged (#41).**
+  A dod block proves one thing — *this story did not break the tree* — and nothing checked that the tree was
+  unbroken to begin with. Measured live 2026-08-31 on `260829-scoring-leaderboard`: of the three commands
+  `workspace.yml` declared, **two already failed on pristine main** — a bare `dotnet test` ran two `Live`-trait
+  tests that call paid Azure AI and that the repo's own CI excludes, and `dotnet format --verify-no-changes`
+  flagged 336 files in a repo whose CI never gates format at all. All 15 stories in the plan would have blocked
+  identically, each having spent a developer turn, and each told the operator the STORY was red.
+  - **Pre-flight at Build entry.** After the dirty-tree and foreign-epic refusals and before anything is
+    dispatched, every dod command the pending stories name is run once against the untouched base tree. A
+    non-zero exit refuses the stage (exit `2`, back to `ready`) naming the command, its exit code and the repo,
+    with no attempt spent and nothing charged.
+  - **In the repo's own checkout, not a fresh worktree.** That is the tree a human means by "the base": it has
+    the installed dependencies and tool state that make the command mean what the team thinks it means. A
+    pristine worktree would fail half the world's repos for want of `node_modules` and turn a safety net into
+    an outage.
+  - **Paid for once.** Results go to `04-build/preflight.yml` — files are the state — keyed by repo, command and
+    the base sha, and are read back by every later invocation of the run. A missing or unreadable cache is a
+    question, never a fault: a run that entered Build on an older binary measures lazily rather than erroring.
+  - **Attribution.** When a story's DoD does go red, the cached base result decides whose fault it is. A command
+    red on the base too halts the build with the same workspace-config error rather than blocking the story and
+    consuming its attempt. A command the gate declined to run is recorded `unmeasured` and excuses nothing.
+
+- **The review handshake no longer swallows an unrecognized verdict, nor drops structured
+  findings.** Measured on `260831-hardening-d1` / S1 (2026-08-31). Two verdict grammars coexist
+  — gate evidence is `sign | sign-with-fixlist | refuse`, a story review is
+  `approve | fixlist | changes` — and the host-facing hint named neither, saying only "write
+  {verdict, summary, findings}". The host wrote `sign`. `parseReview` fail-closed it to `changes`,
+  correctly and **silently**: a clean fix-list verification round read as a second `changes`, the
+  story went `blocked`, and a `story reopen` cycle was the only way to record the verdict that
+  had been meant all along. Separately, `findings` was filtered with `typeof f === "string"`, so
+  the attempt-1 adversarial reviewer's seven `{severity, file, line, claim, evidence, fix}`
+  objects were dropped whole — the verdict survived, the evidence it rested on did not.
+  - **The contract is now stated where the host reads it.** Both `--commit --review` hints name
+    the enum: `verdict is one of approve | fixlist | changes, NOT the `sign`/`refuse` gate
+    vocabulary`.
+  - **The downgrade is announced.** Fail-closed is unchanged — an unreadable verdict is still
+    `changes`, never `approve` — but a verdict outside the enum now comes back on
+    `Review.verdictProblem` ("the reviewer's verdict `sign` is not approve|fixlist|changes —
+    recorded as `changes`"), is printed by the executor on the one path both doors pass through,
+    and rides in `findings` so the review log and the next attempt's `## Previous attempt` both
+    carry it. A DECLARED `fixlist` that fell short is untouched: `fixlistProblems` already says
+    that one out loud, and two sentences for one downgrade would read as two faults.
+  - **Structured findings are rendered, never dropped.** An object becomes
+    `[severity] file:line — claim · evidence: … · fix: …`; a shape nothing recognizes is kept as
+    JSON; a `findings` that is not an array is kept as one finding. An unreadable finding in the
+    log beats a finding that is not in the log.
+
+- **A project stage override that supplies only `stage.yml` no longer swaps the stage body for an
+  empty one.** Reported from the 260829-scoring-leaderboard driver session (2026-08-31) and
+  reproduced here: `stage.md` was resolved by string-substituting `stage.yml` in the path the
+  preset had already picked, so creating `.tldrx/stages/plan/stage.yml` to tune one key moved the
+  BODY lookup into a directory that had none — and the miss was read as an empty string. The
+  context ledger printed `stage 1 B` where it had been 4.9 KB; the sub-agent would have been
+  dispatched with the inputs, the experts and the rejection note and **zero** stage instructions,
+  and nothing refused. `stageMdPath` now resolves per FILE, not per directory: the override's own
+  `stage.md` wins, else the packaged one is inherited, and a stage with no body anywhere is a
+  named `StageBodyError` rather than a silent empty prompt. Both readers — `next --prepare` and
+  the Watch executor — go through it.
+
+- **`approve --as-agent`'s refusal now names the route that works on the run in front of you.**
+  It pointed only at `--gates <stage>:agent`, which is chosen at `run new` and frozen there — so
+  the one suggestion meant recreating a run already in flight. It now leads with the delegated
+  approve: read the agent's evidence note yourself and sign as you, `tldrx approve --note
+  "delegated: <agent> reviewed this, evidence at <path>"`, which keeps the gate's policy and puts
+  the provenance on the record. Found across the 2026-08-30/31 unattended pilots.
+
+- **`budget raise <phase> <usd>` help said `<usd>` was "the new ceiling"; the source adds it.**
+  `raiseBudget` computes `ceiling_usd + amount` (`raiseBudget.ts:83`), so an operator following
+  the help over-raised — measured live on the scavtopia leaderboard run, a "$5.40 new ceiling"
+  command would have set $8.00. The arithmetic is what live runs depend on and is untouched; the
+  words move. `<usd>` is now "how much to ADD to that phase's ceiling — a delta, not a new
+  ceiling", with a note spelling out the $10 + $25 = $35 case and pointing at `budget show`, which
+  already prints the correctly-sized command.
+
+- **`tldrx run new --from` stores an imported answer's own words, not a letter pointing at a file
+  it does not own.** AI-DLC records a chosen option as `[Answer]: C`, and the import stored
+  "<question> — C" verbatim; two facts became unreadable once aidlc was uninstalled and the source
+  file went with it (2026-08-30/31 pilots). The interview flow has always resolved a letter to the
+  option's text before recording (`interview/reply.ts:32-37`), and the import now does the same:
+  `parseAidlcQuestions` reads the lettered options (uppercase, and a space required after the
+  punctuation, so `- E.g. …` stays prose) and `answerText` resolves the answer against them; a
+  letter with no option behind it is stored as typed rather than invented. Conflict detection is
+  unchanged by the longer text: it keys on the QUESTION, as `hooks/no-reask.ts:54` already does,
+  because `findDuplicate` is Jaccard over tokens and therefore length-sensitive — the same
+  contradiction scored 0.78 against a bare letter and ~0.22 against the answer written out.
+
+- **A second run's stories no longer merge into ANOTHER run's epic branch.** Measured live
+  2026-08-31 on two concurrent runs: `260831-hardening-d1` reported S1, S2 and S6 all
+  "merged into `epic/hardening-d1` (N commits carried)", and `epic/hardening-d1` was still
+  sitting at its base with **zero** story commits — all three merges had landed on
+  `epic/d1-tenancy-identity-customers`, a CLOSED previous run's branch. Nothing failed; the
+  run closed green with an empty epic, and it surfaced only because a later story measured
+  `git merge-base` and found the dependency it had been promised was missing.
+  - **The cause was one missing run id.** `openEpicWorktree` built the epic worktree's disk
+    path as `_epic-<epic id>`, and every plan names its first epic `E1`. The second run's
+    `existsSync` therefore hit the FIRST run's directory, `addWorktree` was skipped, and
+    `git merge --no-ff` ran inside a checkout of a foreign epic branch. The in-memory map was
+    keyed correctly (`repo:epicBranch`) — only the path collided, and only across processes.
+    Every progress line renders `story.epicBranch`, so the messages were right about where the
+    merge was *meant* to go for as long as the bytes went somewhere else.
+  - **The path now carries the run**: `_epic-<run id>-<epic id>`, the same shape the STORY
+    worktree was given after the 2026-08-29 audit found the identical class of bug one level
+    down. That fix never reached the epic worktree, which is the worse half — a story worktree
+    collision means two sub-agents editing one file, an epic worktree collision is a merge.
+  - **And a reuse on the wrong branch now refuses.** Every reuse of an epic worktree — the
+    remembered path and the one found on disk — asserts its checked-out branch is the story's
+    epic branch first (`assertWorktreeOn`, `core/build/git.ts`). A mismatch throws
+    `WorktreeBranchMismatchError` naming both branches and the directory, and fails the stage.
+    It never re-points the worktree and never merges anyway. Path scoping makes the collision
+    impossible; this makes it impossible to repeat *silently*.
+
+- **A `--note` with a blank line in it no longer destroys `run.yml`.** Measured 2026-08-31 on the
+  live `260829-scoring-leaderboard` run: `tldrx reject --note "<two paragraphs>"` wrote the note
+  into the gate's flow mapping with LITERAL newlines inside a double-quoted scalar, which is not
+  YAML — the `yaml` package answered `Missing closing " quote at line 57` and Bun's parser
+  `Unexpected character` — and from that moment **every** command on the run failed. There was no
+  repair verb, so the operator had to hand-edit a file the docs forbid editing, and the next save
+  re-emitted the same string and broke it again at the same line, taking `run.yml.bak` with it.
+  Four changes, each closing one part of the loop:
+  - **The emitter escapes, at the one place every field goes through.** `yamlScalar`
+    (`core/facts/emitFactsYaml.ts`) escaped `\` and `"` and nothing else; it now emits via
+    `JSON.stringify`, whose string grammar is a strict subset of YAML 1.2's double-quoted scalar —
+    the escaping this repo already trusted in `adapters/external.ts` and `build/storyFile.ts`.
+    Because every YAML this framework hand-emits routes strings through that one helper, the fix
+    reaches **all** of them at once: gate notes (`approve`, `reject`, `revoke`), `cancelled.note`,
+    task `error` and `stopped_by`, gate `evidence`, run `title`/`scope`, `facts.yml` fact text and
+    retirement reasons, and `split.yml` goals, claims, questions and answers. Verified against
+    **both** parsers behind the runtime seam. Existing files do not churn: over every code point
+    from U+0020 to U+FFFF the new escaping and the old produce identical bytes (63,456 checked, 0
+    differ), so only the values that were already corrupt change shape.
+  - **A file already broken this way heals itself on load.** `parseYamlRepairing` (`core/yaml.ts`)
+    re-escapes raw control characters trapped inside a double-quoted scalar, re-parses, and accepts
+    the result only if it parses — otherwise the parser's ORIGINAL error is thrown, because a
+    repair that cannot be verified is not offered. `RunStore.open` then rewrites the mended file
+    through the fixed emitter and says so on stderr. A one-time hand repair was never enough: the
+    old emitter re-corrupted the file on the next save, so `emit(load(x))` had to be made stable.
+  - **Every state write keeps one step back.** `RunStore` and `FactsStore` had grown a
+    byte-identical private copy of temp-plus-rename each; both now call one
+    `core/fs/writeAtomic.ts`, which additionally copies the version it is about to replace to
+    `<file>.bak`. Atomicity only ever guaranteed a WHOLE file, never a good one. The copy is taken
+    before the rename, so the live file is never absent for an instant and a torn backup can only
+    ever cost a backup. `tldrx init` now adds `tldrx-work/*/*.bak` and `.tldrx/memory/*.bak` to the
+    managed `.gitignore` block.
+  - **A `run.yml` beyond mechanical repair fails honestly, and takes nothing else down with it.**
+    The error names the file, quotes the parser verbatim, says that every command on the run reads
+    that file first, and points at `run.yml.bak` — while stating plainly that using it is a MANUAL
+    decision tldrx will not make. Separately, one corrupt `run.yml` used to throw a raw
+    `YAMLParseError` out of `buildModel` and kill `tldrx dashboard` for the whole workspace;
+    `loadRunResult` (`core/replay/loadRun.ts`) now distinguishes missing from unreadable, and the
+    dashboard lists the run as **unreadable** with the parse error beside it and renders every
+    other run as normal.
+
+### Changed
+
+- **The docs now say, at the top of both places a reader starts, that `run attend host` is a LOCK
+  and `run auto` is an ENGINE.** Grounding: on 2026-08-31 the framework's own author — who had read
+  the chapter — ran `tldrx run attend host <run>` expecting it to drive the whole run by itself,
+  and then asked whether `attend` and `auto` compose. They do not, and the code has always said so
+  (`runAuto.ts:108` refuses `run auto` on an attended run at exit `1`, before the event log is
+  opened; `runNext.ts:659` exits `4` on a bare `next` and names the `--prepare` command). The docs
+  took too long to say it.
+  - **README gains "Trying it: three ways to run"**, immediately after Quick start: a three-row
+    table of who executes each turn, what a turn costs, and where each mode stops; one scenario
+    line each (`run auto` for a small run you would watch anyway and for CI/cron — the only mode
+    with no session behind it; `attend host` when a session is already open and cost or quality
+    matters; `attend host` + a mandate for overnight); and the two-command recipe with a verbatim
+    example **mandate prompt**.
+  - **`docs/guide/10-unattended-mode.md` leads with the same disambiguation** — a blockquoted
+    lock-vs-engine table above the chapter's opening paragraph, so a skimmer cannot make that
+    mistake — and gains a `### The mandate` section carrying the prompt verbatim, tying its four
+    legitimate interrupts back to the `questions` / `budget-event` / `boundary` fallthroughs the
+    framework already enforces, and to the fact that no `git push` wrapper exists in the Build
+    executor (`src/core/build/git.ts:13`) and the developer prompt says "Do not push"
+    (`src/core/build/prompts.ts:180`).
+  - **`tldrx run --help` says it too.** `run` had notes for `status`, `estimate`, `unlock` and
+    `cancel` and none for the pair that actually confuses people. It now leads with one note per
+    mode — "a LOCK, not an engine" / "an ENGINE, not a lock", each naming the other's refusal — and
+    the `<host|--none>` argument line says the framework will not spawn on the run again. Help text
+    only; no behaviour, no flag and no exit code moved.
+
+- **`tldrx run estimate` is remaining-work aware (#21).** It priced the next stage from token medians while
+  the budget brake separately computed what that stage still had to pay for — two models, one question, and
+  the one people read was the one that never shrank. A Build stage with five of six stories done was still
+  quoted the number the Plan wrote before any of them ran, which is the figure that made a pilot operator
+  move money twice for work the run could already afford.
+  - It now calls **the same `remainingWork()`** the brake and `budget show`'s `est.` column call, with the
+    same inputs, and reports it beside the token estimate: done stories excluded, blocked ones named, the
+    arithmetic shown. A test asserts the two numbers are identical rather than merely similar.
+  - It also rolls the run up: `still to run: N stage(s) … $X priced`, with terminal stages excluded and the
+    cursor stage narrowed by the plan when the plan knew better. `--json` carries both as `remaining` and
+    `runRemaining`.
+  - The token half is untouched. The input side is still measured off the same assembly `next` builds, and
+    the cache/output medians still say which sample they came from — that half was never the complaint.
+
 ## 0.3.1 — 2026-08-31
 
 **Unattended mode.** Twelve of the entries below are one feature: a run a **host session**
@@ -653,6 +2976,66 @@ none of these behaves byte-identically to the release before them.
 
 ### Fixed
 
+- **A trainer that `cd`s no longer writes its knowledge file into a different git repo.**
+  Measured 2026-08-31 on `~/scavtopia` (five repos, ten `expert train --mode light` runs): the
+  `mcp` run was rejected with `mcp.md.partial was never written`, and the file had been written —
+  46 lines, 9,567 bytes, complete and usable — to
+  `whiteboard/.tldrx/experts/mcp/knowledge/mcp.md.partial`. The sub-agent ran
+  `cd <workspace>/whiteboard` to execute that repo's declared gate command, then wrote the
+  RELATIVE path the prompt had given it, and the path resolved against the repo it had `cd`'d
+  into. Three costs from the one bug: **$1.23 charged for work that was finished and then
+  orphaned**, a parasitic `.tldrx/` tree left inside an unrelated git repo (`git -C whiteboard
+  status` → `?? .tldrx/`), and **no repair round possible** — the missing-file branch returns
+  before the repair check, so this failure mode was unrecoverable by construction even with
+  budget left. Fixed at both ends.
+  - **Prevention: the prompt now states an ABSOLUTE output path**, workspace-root-resolved, and
+    says why — "If you `cd` into a repo to run its gate command, a relative path then resolves
+    against THAT repo … and throws the whole paid run away. That is measured, not hypothetical."
+    Both training prompts carry it, and so does the repair round's target, for the same reason.
+  - **Recovery: when the file is missing, the declared repo roots are probed** for the stray
+    relative write before "never written" is said. A file found there is moved back and validated
+    exactly as if it had landed correctly — recovery is not a pass, the same `parseKnowledgeFile`
+    still judges it, and a recovered file that fails can still be repaired because the probe runs
+    ABOVE the repair round.
+  - **The note is honest and names the mess.** `recovered: the trainer wrote to
+    whiteboard/.tldrx/… , inside the `whiteboard` repo — a relative `.tldrx/…` path resolves
+    against whatever directory it had `cd`'d into.` The empty parasitic directories are removed
+    on the way out; a directory holding anything else is **left in place and named**, with the
+    `git -C <repo> status` to run, because a tool that deletes inside a repo it was never asked
+    to touch is a worse bug than the one it is fixing. A repo carrying its own
+    `.tldrx/workspace.yml` is skipped entirely — that file may belong to a nested workspace, and
+    taking it would be theft rather than recovery.
+  - When no stray is found the verdict is unchanged and now says where it looked.
+
+- **A rejected training run records WHICH problems, not just how many.**
+  Measured 2026-08-31: `components` failed with 12 problems for $1.02, and `training.jsonl` — the
+  durable record — held only the string `"…does not validate — 12 problem(s)"`. The twelve went
+  to stdout, where five of them were printed and the rest elided as `(+7 more)`. Anyone who had
+  not captured stdout, which is anyone running this normally, could not tell why a $1.02 run
+  failed. The list is now persisted twice.
+  - **On the ledger**: `check.failed.payload` carries `problems` (the rendered per-problem
+    lines), `problems_total`, `errors`, and `task`. The list is fitted to the record's 4 KB
+    payload cap and reports `problems_omitted` when it does not fit — an append that THROWS on
+    an oversize payload would take the cost line down with the reasons, which is the opposite of
+    the point. The repair round's own `check.failed` carries what it sent back, so "what did the
+    repair actually fix" is answerable later.
+  - **In the file**: `<area>.rejected.md` now opens with a `# REJECTED` header — expert/area,
+    mode, timestamp, dollars spent, error and warning counts, and every problem, uncapped —
+    above the trainer's bytes exactly as written, separated by a rule. A quarantine with no
+    verdict (a sub-agent that died, a rollback) gets no header: there were no reasons to state
+    and inventing them would be inventing the reason.
+
+- **`## Sources` is now taught as prose with the refused shape shown.** Same batch: four of the
+  five problems the `components` report printed are one mistake four times — `L34 Sources: no
+  [src: …] token`, `L35`, `L36`, `L37`. The trainer had written the recap as a bulleted list.
+  The prompt already said "**Sources** — prose", and a writer who reads that as a style note
+  writes bullets, because bullets are what the other four sections take. The rule it collides
+  with is genuinely file-wide — `parseKnowledgeFile` requires a `[src: …]` token on EVERY list
+  item in every declared section, recap included, and an unsourced one is an error that rejects
+  the file whole. Both prompts now show the accepted prose next to the refused bullets, the same
+  move the execution-claim rule makes. Whether an unsourced recap bullet should be a warning
+  rather than an error is a real question and is deliberately NOT settled here.
+
 - **A rejected knowledge file gets ONE repair round before the money is thrown away.**
   Measured 2026-08-30 on `~/scavtopia`: `tldrx expert train dotnet-stack --area dotnet --mode
   light` spent **$1.69**, the trainer wrote `knowledge/dotnet.md.partial`, and the validator
@@ -985,6 +3368,19 @@ none of these behaves byte-identically to the release before them.
 
 ### Verified, not changed
 
+- **`tldrx expert train` already exits nonzero when a training fails.** The 2026-08-31 batch
+  report measured shell `EXIT=0` on all ten invocations, including the three that failed their
+  check — but nine of those ten ran on a build that predates this one (`dist/tldrx.js` was
+  rewritten mid-batch at 05:40Z). On the current source the code path is intact:
+  `runTraining` returns `EXIT_AGENT_FAILED` (5), `expert train` returns `outcome.code`,
+  `dispatch` returns it, and `bin/tldrx.ts` does `process.exit(await dispatch(...))`. Now pinned
+  by three tests that drive the REAL CLI as a subprocess with a fake `claude` on PATH and assert
+  the PROCESS exit code — one for a file that does not validate, one for a file that was never
+  written, one for the passing case — because "`runTraining` returns 5" and "the process exits
+  5" are two different claims. Falsified before being trusted: making `expert train` return
+  `EXIT_OK` breaks two of the three.
+
+
 - **The walk already skips vendored and generated trees**, and always did: `SKIPPED_DIRS`
   in `detect/walk.ts` covers `node_modules`, `dist`, `build`, `out`, `bin`, `obj`,
   `target`, `.venv`, `Pods`, `.next`, `.expo`, `coverage` and more, plus every