bareloop 0.10.0 → 0.12.0

package/CHANGELOG.md

@@ -5,6 +5,516 @@ All notable changes to bareloop are documented here. Format:
 [SemVer](https://semver.org/spec/v2.0.0.html). Pre-1.0: **minor** = a ladder rung or
 feature lands, **patch** = docs, fixes, scaffolding.
 
+## [0.12.0] — 2026-08-23
+
+### Added (N4 slice 2 — the softgreen rung: a close that JUDGES, and a signer who has the last word)
+
+- **The judged floor is real: LOCATE + DECIDE (`src/judged.js`).** A judge (haiku-4.5, pinned —
+  never a step knob) extracts FACTS from the real artifact with quotes; an arbiter-owned `decide()`
+  renders pass/fail from a signed rubric card; *unsure* is a red, never a pass. The judge never
+  renders a verdict, so the rubric-is-self-consistency gotcha stays closed. Wired as the
+  `judged-floor` close kind — metered from the run's own wallet, `offer:false` by law, and skipping
+  the seed-verdict read (its bar comes from calibration instead).
+- **`soft-green` is ADMITTED as a verdict class** — the class-scoped question set, the inherited
+  green guard battery (hamr: *"battery ties with job type"*), the class-scoped kind menu, and the
+  promise rule: composition stays ≤ the picked class, and exceeding it is an honest red rather than
+  a silent up- or downgrade.
+- **The signer's two artifacts are AUTHORED and enumerated in the hash** (`src/cardauthor.js`): the
+  rubric card comes from Q6 and the calibration set from Q7, both handed to the compiler verbatim,
+  both SHOWN to the signer (D5 shown-and-fixed), and what the signer fixed is what is stored —
+  `source: 'proposal'` says *nobody changed this*, which is a different fact from *a person read it
+  and changed nothing*.
+- **The calibration gate — ten graded truths and five resisted lies buy a signature**
+  (`src/calibrate.js`, gate 4 of `prepareSigning`). The whole pipe grades the signer's own ten cases
+  and faces a 5-artifact injection battery adapted to the locate axis; the floor is **10/10 with
+  itemized reds** (hamr, 2026-08-18) and all-or-nothing on both halves. The $0 legality check runs
+  first and buys no calls. A dead judge is a CASUALTY under the transport's name, never a reading of
+  the operator's rubric.
+- **Quarantine: a judged green earns no learning credit until an `accept` releases it.** Minted
+  soft-green bridge rows are held at mint, the release is forward-only, a rerun re-quarantines, and
+  a held entry is never offered by selection — so a soft-green pass cannot feed N3 before the judged
+  floor is proven.
+- **The review door — every run ends at the signer's three doors** (`accept` / `rerun` / `pause`).
+  The door records a DISPOSITION and never touches the loop's own verdict (hamr: *"it's important
+  not to change the loop self verdict"*): `runPlan` returns `green`/`already-green` exactly as
+  before. It opens at all three green sites (precheck `already-green`, the outer close, and the fix
+  loop's green). **Soft-green always** (`REVIEW_DOOR_CLASSES`), **green opt-in** via `reviewDoor:
+  true` / `--review-door` — the flag wins in both directions and is never inferred. An `accept`
+  re-runs the close's MECHANICAL stages (never a judged floor, never a person): a door keeps for 60
+  days and nothing freezes a tree.
+
+- **The judge tier joins the SIGNATURE — a bump flips the hash and the stage refuses the
+  mismatch** (`649d8a5`). `JUDGE_MODEL`'s docstring has always said a judge-model bump forces a
+  full recalibration, and nothing detected one: `closeDecl.calibration` stored only its `cases`,
+  so the tier that graded them sat OUTSIDE the signed bytes, a bump flipped no hash, and the
+  judged stage never compared the model it was about to buy against the model the signer's set
+  was graded by. A rule with no wired detector is prose (F45). The tier now lives beside the
+  cases as **`calibration.judgeModel`** (`CALIBRATION_FIELDS` in `src/declaredclose.js`
+  enumerates a stored set as exactly those two and nothing else), written by
+  `foldJudgedArtifacts` from the PIN itself rather than from anything a caller may name — and
+  because it is inside the spec, `jobSpecHash` covers it by construction: bump the constant,
+  every signed judged spec hashes differently, the signature dies, and the re-sign is what
+  re-runs the calibration gate. `validateCloseDecl` REQUIRES it on any stored set (a set nobody
+  can attribute to a tier is a floor nobody can tell apart from a bumped one).
+  `declaredStages` carries the close-level fact onto the judged stage as
+  `calibrationJudgeModel` — the same seam, and for the same reason, that already stamps
+  `offer: false` — and it carries it from the SIGNED bytes, so a stage can never name a tier the
+  signature does not. `runJudgedFloor` then refuses to grade on a mismatch, naming both models
+  and the two things that fix it (*re-sign* and *recalibrate*), under the same `FAILED` fault the
+  absent judge seam already stops on: nothing is broken, something was never re-done. An ABSENT
+  stamp is deliberately NOT a runtime stop — a signed spec cannot reach a run without a
+  calibration (the signing gate) or without a model on it (the spec gate), so the only runtime
+  question is WHICH model. No signed spec on disk moves: `jobs/*.json` carries zero
+  `judged-floor` stages today.
+- **The U runner mints the row its review door promises to release** (`e01d0af`). The door tells
+  a person that `accept` *"releases this run's learning credit"*; for every run
+  `scripts/run-u.mjs` has driven that sentence was false by construction — a cold green wrote a
+  standalone `bridge-<job>-<runid>.json` FILE and never a registry ROW, so
+  `--registry --workflow --decide accept` reached `recordDoor`, found no green row for the runid
+  and died `no-row-for-run`. The held credit could be described and never released. **STORAGE
+  ONLY:** nothing here selects, promotes or reuses a bridge — that rung stays parked on
+  `layer-3-reuse` (F88) — and exactly one thing reads the row back, the door. `writeGreenRow` is
+  `runReuse`'s own cold-leg green write LIFTED out of its closure unchanged (the
+  fork-on-a-different-close-shape rule, both collision refusals, the `appendGreen` leg), so the
+  two runners mint a green through ONE spelling; a second spelling of what a green writes is the
+  two-transforms class applied to the ledger itself. **`writeRunGreenRow`** (newly exported) is
+  the U runner's terminal policy on top of it, and every guard is a NAMED refusal rather than a
+  silence: no `--registry` (`no-registry`), an already-green run (`green-predates-run` — accept
+  confirms a verdict, it never mints one), a green with no executed plan (`no-plan-executed`), a
+  class whose credit is not held (`credit-not-held` — `green`-class runs mint nothing here,
+  which is what this runner has always done), and a spine leak blocks the row exactly as it
+  already blocks the bridge file. The row is born HELD because `greenParts` reads the SIGNED
+  class off the record — nothing new sets a flag — and `applyDoorDecision` releases it through
+  the path that already existed. `src/index.js`'s claim that `runReuse` is the only registry
+  writer is CORRECTED rather than left standing.
+### Fixed
+
+- **F102 — a pending human decision now survives a wall-halt → resume** (`a6f62c9`). Two records
+  answering two different questions: `human-decision` (the door, the words, and by whom, with
+  `source: 'operator'|'checkpoint'` and the moment the PERSON spoke) and `human-decision-spent`
+  (what actually bought it). The spend marker fires on the **WORK** — a fix round bought, an accept
+  that greened, a pause honoured — **not on the fix loop opening**, which is precisely the shape of
+  the incident (`iterationsUsed: 0`). A resume carrying one re-enters the fix loop through the same
+  seam a fresh answer uses; the ask is never re-rendered. Two answers to one question are REFUSED
+  (`hitl-decision-red` naming what is held), never merged.
+- **F103 — a rerun buys its own clock** (`a6f62c9`). Two counters that never do each other's job:
+  CHAIN (`chainWallMs`/`chainSpentUsd`, `job-end.spentUsd`) is every leg added up and is what the
+  halt readout reports; ENGAGEMENT (`priorElapsedMs`, `job-end.engagementSpentUsd`) is what bounds
+  THIS leg. Money is deliberately not symmetric — the signed budget stays the CHAIN's ceiling,
+  because a refilled wallet is the failure F45 measured. W-2 is untouched where it was designed for:
+  an ordinary resume folds exactly as it always has, and so does an `accept`.
+- **F106 (HIGH, review) — the calibration gate now meets the one ceiling** (`0ae7693`). The run's
+  third and largest paid seam (up to 10 locate calls plus the 5-artifact battery, each under its own
+  retry ladder) ran with no bound while the scout and the declaration loop were both stopped by
+  `run-author --budget` — the advertised budget and the enforced budget were two different numbers.
+  `prepareSigning`/`calibrationGate` take `ceilingUsd` + `priorCalls`, a cost book absorbs prior
+  spend so the gate reads the REMAINING balance, `capStop` is asked before each paid call inside the
+  pipe (retries included), and a ceiling stop returns as an operator refusal naming money
+  (`cap-halt` / `pricing-red`) — never a casualty, and never a part-graded set presented as graded.
+- **A review door's rerun is a fresh engagement too** (`0ae7693`): `resolveHumanRuling` takes
+  `doorRerun` as a third input, OR'd and never merged, so *"is this leg a fresh engagement"* keeps
+  ONE spelling and a door-rerun leg cannot inherit the rejected leg's wall.
+- **`scripts/run-u.mjs` wires `judgeProvider`** (pinned to the judge model, never a step knob) — a
+  live soft-green run no longer instrument-stops at the judged stage.
+- **Two mutation-blind gaps in `src/kinds.js` covered by tests** (found by the same review).
+
+- **The medium whole-branch code review's five findings** (`e0c87d3`). An absent `job-end` in
+  `scripts/run-u.mjs` (a crash between the door record and `job-end`, or a legacy spine) no
+  longer launders into an exact complete `$0` that folds into the rerun ceiling — unknown is
+  reported unknown (F6), regression test watched failing first. One name per kind and per code:
+  the `'judged-floor'` literal at 7 sites becomes **`JUDGED_FLOOR_KIND`** (`src/kinds.js`,
+  exported) and the `'quarantined'` code at 3 sites becomes **`QUARANTINED_CODE`**
+  (`src/bridges.js` — deliberately not `reuse.js`, which would cycle reuse ↔ selection). One
+  `threeDoors()` builder replaces the copy-pasted door lines in `scripts/u-readout.mjs` (output
+  byte-identical against HEAD), and a `changedEvidence()` helper in `src/planrun.js` replaces
+  the duplicated evidence block and the double `verdictType` default, with `doorOpens`'
+  signature untouched. Two findings were ESCALATED rather than decided — the judge-model bump
+  with no detector, and the U runner's door promising to release credit it never minted — and
+  both are the `1A`/`2B` entries above.
+- **The tail review's six findings over `1A`+`2B`** (`493b2ce`). The door PREVIEW path now
+  checks row-EXISTS before promising a credit release, and its printed commands carry
+  `--registry`/`--workflow` forward — the `e01d0af` fix had landed at the end-of-run readout and
+  not at this sibling site, so the operator's next invocation would have rediscovered
+  `no-row-for-run` one hop later. The registry row records **`rounds` + `roundsComplete`**: a
+  resume leg writes a fresh spine and the halted leg's restart record declares money and wall
+  but never rounds, so on a resume that count is a FLOOR and now says so the way the money
+  beside it already does (F6). A NAMED `--registry` that attempted no row prints the reason
+  rather than staying silent. The two near-identical record builders in `src/reuse.js`
+  consolidate into `commitWrite`, and the duplicated `worker-round` scan is hoisted to one
+  spelling.
+- **The release gate's two findings** (`a5a4731`). `docs/UPSTREAM-ASKS.md`'s BA-21 carries a
+  dated CORRECTION note: the retracted 2.0–2.5× figure was a clipipe-surface number (F48 — a
+  provider-billed surface that never pools with `anthropic-api`), and the measured API-side
+  numbers are ~5.7% / 1.586×; the acceptance criteria are unaffected. In `src/judged.js`, two
+  literal NUL bytes inside composite-key template strings are spelled `\x00` — byte-identical
+  at runtime, but the literal made `grep`/`file` treat the source as binary (the documented git
+  binary-detector trap).
+
+### Docs (2026-08-23)
+
+- **`docs/FINDINGS.md`: F102 and F103 move from NOT FIXED / NOT BUILT to FIXED**, each cited to
+  the commit that did it — `a6f62c9` for both, plus `0ae7693` for F103's extension (the review
+  door's rerun as `resolveHumanRuling`'s third input, OR'd and never merged). **F106 is minted**
+  from the 3-lens review's HIGH: F45's class in a money coat with the direction reversed — the
+  ENFORCED number was the looser one, which is the failure that leaves a bill instead of a red
+  (`cce3ce5`).
+- **PRD Addendum v1.76** (this release): the three doors restated, the green-class review door
+  kept OPT-IN, rerun named as *not a replay*, the `1A`+`2B` escalations ruled, and the one
+  remaining door hole named and left deliberately.
+- **`bareloop.context.md`**: the judge-tier paragraph rewritten from DEFERRED to the live
+  detector it now describes.
+
+## [0.11.0] — 2026-08-18
+
+### Changed (2026-08-18 — the review door's third button)
+
+- **The three doors are `accept` / `rerun` / `pause`; `cancel` is DELETED as a concept.** hamr:
+  *"what's the point of cancel anyways? pause can resume — that would be more honest"* and
+  *"rerun implies I don't like it, go again"*. A `pause` answer runs nothing, spends nothing,
+  buys no worker round, never opens the fix loop and does NOT consume the one-shot rerun
+  allowance: it mints the ordinary `hitl-pause` checkpoint with an explicit
+  `humanDecision: 'pause'` and leaves the run resumable from the start of its last step. An
+  unresumed pause simply expires under the existing 60-day `PAUSE_TTL_MS` — **that expiry is
+  what cancel used to be**, without forcing a permanent decision. `HUMAN_DECISIONS` is the
+  enumerated set the surface derives from, so `cancel` is now inexpressible rather than
+  rejected after the fact; `hitl-decision-red` still refuses any word that is not a door.
+- **`hitl-cancel` is no longer mintable and the constant is gone** (removed from
+  `src/declaredclose.js` and the package exports). The ledger keeps the bare string in its
+  excluded-escalation set — the `gate-red` precedent — so a spine written before this change
+  still reads as governance rather than as a counted capability gap; no spine on disk carries
+  it (the door never fired live), so that entry is forward compatibility, not bookkeeping.
+- **A human checkpoint row is classed `checkpoint`, not `casualty`** (`src/reuse.js`, both the
+  live try loop and the resume reconstruction). A casualty is a run that died; a checkpoint is
+  a run that is waiting on a person with nothing lost — every other reader already treated it
+  that way, and the row was the last place the machinery mis-described itself.
+- **Runner surface (`scripts/` never ships):** `scripts/run-u.mjs --decide pause` answers the
+  door and launches nothing at all — it prints the decision, names the TTL and the ORIGINAL
+  runid, and exits having spent nothing. Deliberate: the runner writes one spine per leg, and a
+  leg that returns before drafting emits no `plan-accepted` and no `step-end`, so launching a
+  run to say "not now" would mint a runid whose own checkpoint is empty and re-pay for finished
+  steps on the next resume. Door rendering (`scripts/u-readout.mjs`) and the escalation
+  passthrough (`src/ralph.js`) carry the new third door.
+- **Docs:** dated addendum on `docs/02-features/2026-08-17-softgreen-review-door-design.md`
+  (closed sections untouched), **PRD Addendum v1.72**, README's *"You have the last word"*,
+  `bareloop.context.md`'s terminal/decision surface, and `docs/01-product/LAYERS.md`.
+
+### Docs (2026-08-17 — decisions only; NO code changed, and no build is authorized)
+
+- **New design record `docs/02-features/2026-08-17-softgreen-review-door-design.md`, and PRD
+  Addendum v1.71: `hitl` is RETIRED as a verdict class, its pause becomes the REVIEW DOOR at
+  the end of EVERY run, and `soft-green` becomes the forward path.** The hitl slice below FIRED
+  live first and worked — the first hitl pause in programme history, 64 strict errors to 0, the
+  authored guard catching the cheat genre twice, evidence package and three doors and a stopped
+  clock (F105). hamr then retired the class for what it asks of a composer and of a person:
+  *"checkers are subjective human experience that they should grade for"*, and *"it's hard to
+  apply deterministic flow on a probabilistic throughput"*. **The machinery is re-homed, not
+  deleted.** The door never changes the loop's own verdict (*"it's important not to change the
+  loop self verdict"*): a green stays green in the ledger and the door records a disposition —
+  `accept` (and, on soft-green, the release of quarantined learning credit), `rerun` (the text
+  IS the gap), `cancel` (*"the verdict stands, but I'm not taking the merchandise"*). A rerun is
+  a **FRESH ENGAGEMENT** with its own money and time and two side-by-side counters, and **any**
+  wall or halt now records full state — time, money, strikes, notes/progress and any pending
+  human decision. `soft-green` gets a `judged-floor` kind on the LOCATE+DECIDE pattern (the
+  judge extracts facts with quotes from the real artifact; an arbiter-owned `decide()` renders
+  pass/fail; unsure = red), behind a signed rubric card (Q6) and a signed calibration set (Q7)
+  the whole pipe must grade correctly before the close is signable. **Nothing is built.**
+- **F102, F103, F104 and F105 in `docs/FINDINGS.md`**, each grounded in today's spines under
+  `bareloop-patients/litectx-maintainer-bareloop/`: a pending rerun decision does not survive a
+  wall-halt → resume (two byte-identical pause records — the human paid the same decision
+  twice); the wall folds across legs, so a decide-time rerun inherited 87 seconds; the
+  catalogue cannot count documentation, so a composer measured a typecheck PROXY instead —
+  which is what retired the class; and the positive record of everything the first live hitl
+  loop proved, including that rerun DELIVERY works and rerun CONVERSION is still unproven-live.
+- **`docs/01-product/LAYERS.md`** gains the review door, the two-verdict table and the next
+  rung; the **2026-08-13 N4 build record** (closed) gains a dated addendum pointing forward;
+  **README** and **`bareloop.context.md`** carry the direction change, the contract still
+  describing what the code on this branch actually does today.
+
+### Added
+
+- **N4 slice 1 — the `hitl` verdict class is ADMITTED, and the run PAUSES for a person.** A job
+  whose *done* needs someone to say so now runs through the same signed spec, the same
+  agent-authored plan and the same arbiter as a green job. The refusal was never one flag, so
+  every coupled site moved together: `LOCKED_VERDICTS` / `LOCKED_CLASSES` drop it,
+  `DECLARED_CLOSE_CLASSES` widens to `['hard','hitl']`, the class battery is the GREEN mechanical
+  battery inherited verbatim (hamr's OPEN-1 ruling: nothing a human stage cannot see), and the
+  interview gains a hitl question set (the green questions byte for byte, plus the one thing
+  they cannot supply — the question the signer actually answers). **`soft-green` stays
+  declared-but-locked**: its judged floor is slice 2, and a wobbly ruler must mint nothing.
+- **`human-confirms` goes live as the one kind that measures nothing.** Its whole parameter
+  surface is `ask`; it cannot spawn, is never env-capable, is `offer: false` BY LAW (a
+  declaration that offers one is a validation red, and the arbiter's own bridge stamps the law
+  onto the stage the runner sees, so the agent can never compose `check-passes(<a person>)`), is
+  at-most-once, and must be the LAST stage — first-red-wins would otherwise silently delete every
+  mechanical stage behind it. It also SKIPS the seed-verdict read (ruling 8), and the skip is
+  recorded as a row rather than taken in silence.
+- **Three new terminals, minted rather than borrowed.** `hitl-pause` is a decision-ready
+  CHECKPOINT carrying the evidence package (every mechanical stage's result, the close's own
+  question, and what the run changed) — never a bare "approve?"; `hitl-cancel` is terminal with
+  an explicit `gap: null`, no fix loop and no worker round; `hitl-decision-red` refuses a
+  decision the run cannot act on (a fourth door, a `rerun` with empty text, or a decision handed
+  to a close with no human stage) before anything is spent. All three are excluded from the
+  ledger's escalation counting for their own stated reasons, and none demotes a bridge. The
+  legacy `hitl-close` entry keeps its own meaning; nothing was renamed in passing.
+- **The RETURN needs no new channel.** A resumed leg carrying the signer's `rerun` feeds their
+  words through the same seam `post.gap` uses — same bound, same scrub — and the run continues
+  under the folded wallet on an unbilled clock. The ruling is spent when the fix loop opens, so
+  the next machine-clean tree pauses for a second review instead of converting one sentence
+  forever, while a mechanical red in between is still converted normally.
+- **New public exports**: `HUMAN_DECISIONS`, `SEED_EXEMPT_KINDS`, `normalizeHumanRuling`,
+  `NEVER_OFFERED_KINDS`, `HUMAN_PAUSE`, `CHECKPOINT_OUTCOMES`, `PAUSE_TTL_MS`,
+  `checkpointAgeGate`, `HITL_PAUSE`, `HITL_DECISION_RED`, `HUMAN_CHECKPOINTS`. The 60-day pause TTL and the
+  canonical checkpoint list live in the LIBRARY (hamr's OPEN-2 ruling) so the exported bundle
+  inherits them rather than every runner re-implementing them. The two terminal names are
+  spelled ONCE, in `src/declaredclose.js` beside the `HUMAN_PAUSE` close-verdict word they
+  translate, because four different readers key on them — the emitters (`src/planrun.js`), the
+  ledger's excluded-escalation set, the resume reader (`src/reuse.js`) and the runner script —
+  and hand-spelled copies of one terminal is how those readers come to disagree.
+- **A registry guard that was only holding by accident**: an `already-green` terminal now mints
+  no bridge version for a NAMED reason (`green-predates-run`) rather than because no plan
+  happened to be on the spine — a hitl try that paused and came back with `accept` has its
+  predecessor's plan in the very same try window.
+- **An authoring run says what it is doing while it is still doing it — three REPORTING seams,
+  `onPhase` / `onStage` / `onCall`.** The pipeline ran up to ~15 minutes between `author-start`
+  and its result with nothing on the terminal or the spine — a real survey ladder, a real
+  declaration ladder, and a real toolchain per close stage, all inside one await — and silence
+  is byte-for-byte what a hang looks like, so the operator's only lever was to kill a run that
+  might be working. The library REPORTS and the shell PRINTS, the same shape `runJob` → `runPlan`
+  already has: `onPhase` fires at the composition's boundaries in `authorCloseForJob`
+  (`src/authorjob.js`) and `authorClose` (`src/authorflow.js`) — `seed`, `scout`/`scout-done`,
+  `listing`/`listing-done`, `author`, `author-call`, `seed-read`/`seed-read-done`; `onStage`
+  (a new `Ctx` field in `src/kinds.js`) fires per seed-read stage, AFTER the stage lands, because
+  only then is there a verdict to carry; `onCall` fires from inside each of the two metered
+  recorders (`makeCostBook.add`, `runAuthorScout`'s `record`) rather than at the call sites, so a
+  paid call cannot be metered without its report going out. **Nothing here is consulted and
+  nothing decides**: `capStop` is still the one ceiling predicate, still asked between calls,
+  still reading the same entries — and every callback defaults to a no-op, so every existing
+  caller is byte-identical. Two deliberate omissions: `makeCostBook.absorb` does NOT fire
+  `onCall` (those calls were already metered and reported by whoever made them, and reporting
+  them twice would double the scout's spend in every reader downstream), and the `seed` phase is
+  announced only on the path that actually reads one, because a phase line for work nobody did is
+  the blind-instrument class one line wide. The renderer is `phaseLine` in
+  `scripts/author-readout.mjs` (a readout no test can reach is a readout nothing checks); it
+  prints WHAT is happening and never a fraction — a survey attempt has no progress and a suite
+  stage finishes when it finishes — and an unmeasured `durationMs` prints as `unknown`, never as
+  `0` (F6 in its time form). A phase the renderer does not know is printed rather than swallowed.
+
+The runner half of the same slice landed beside it and is deliberately NOT in the tarball
+(`scripts/` never ships): `scripts/run-u.mjs` gains `--decide accept|rerun|cancel` (+ `--text`)
+on the existing `--approve <specHash>` signature, renders the pause's evidence package plus the
+line-level diff, consumes `CHECKPOINT_OUTCOMES` and `checkpointAgeGate` instead of re-spelling
+either, and prefers the run's own terminal over a stale watchdog record when folding the wall —
+so a person's deciding time is never billed to it. What an adopter's own runner has to do is in
+`bareloop.context.md`.
+
+Beside it, `scripts/run-interview.mjs` — **the terminal interview, and the answer to hand-writing
+an `answers.json`**, which is precisely the SWE tax this product exists to refuse. It asks the
+LIBRARY's frozen questions ONE AT A TIME, **calls no model and costs $0**, and writes
+`answers.json` (the shape `run-author.mjs` consumes) plus `specdraft.json` (the operator half — no
+close and no `verdictType`, because those are what `run-author` authors). The draft is checked for
+$0 by `validateJob`, the SAME validator that will judge the resolved spec, so a draft that cannot
+become a spec says so before a provider is ever built. It then OFFERS to spawn the paid
+`run-author.mjs` under its OWN `--budget` ceiling — **default NO** (`[y/N]`, the same lean the
+pause's three doors take: the answer that costs nothing is the one you get by saying nothing), and
+it prints the exact command either way. The two ceilings are kept apart out loud: `--budget` is
+what the AUTHORING call may spend, `budgetUsd` is what the RUN may spend, signed into the spec.
+
+And `scripts/run-u.mjs` gains a `litectx-maintainer` JOBS row — **N4's hitl PROVING job, dark
+since the legacy `steps[]` deletion took its old spec's path with it**, back as a plan-flow job
+with an AUTHORED close. `jobs/litectx-maintainer.json` **DOES NOT EXIST, and the row does not
+create it**: the spec is authored live through `run-interview.mjs` → `run-author.mjs` and signed
+by hamr, goal and answers included, and until it lands **the runner REFUSES BY NAME** — a row is
+the runner's half of a job, never the signer's. The patient is a COPY at the convention this table
+already uses, and the seed is that copy's own HEAD (litectx v0.32.0, `115213d`).
+
+### Changed
+
+- **Both suite pins move up a minor — `bare-agent ^0.35.0 → ^0.36.0`, `bareguard ^0.12.0 →
+  ^0.13.0` — because on 0.x a caret locks the MINOR**, so the old pins admitted neither release
+  and the bump is the only way to consume them. **bare-agent 0.36 delivers BA-20**, the
+  productized decisive judge this queue reopened on: `judge` (verbatim request + one structured
+  artifact → `honored`/`broke` with a mechanical `where`, truncated/unparseable responses as
+  DISTINCT flagged outcomes excluded from graded denominators, composing **around** a provider
+  and never inside the Loop), the calibration harness (`calibrate`, `CALIBRATION_CASES`,
+  `INJECTION_BATTERY`, `constantHonored` — the negative control that lets the harness fail), and
+  the pure `judgeToAnnotation` mapping into bareguard's shipped fact envelope. **bareguard
+  0.13.0** is the companion hardening: `gate.annotate` now REJECTS malformed facts into distinct
+  `annotate_malformed` rows instead of normalizing them into silently-`honored` ones, `verdict`
+  goes through the audit redactor (it was written RAW — a real secret-leak fix), `meta`'s
+  1000-byte bound became a decoupled copy, and `__proto__` keys are dropped at every depth.
+  **That behaviour change is zero-cost to bareloop today**: it imports only `Gate` and `redact`
+  and has **no `gate.annotate` call site** at all. `npm ls` resolves `bare-agent@0.36.1` +
+  `bareguard@0.13.0`, deduped to one shared copy, and **full fresh gates are green against the
+  installed versions** — `npm test` 1569/1569 exit 0, `typecheck` clean, `build:types` clean.
+  **The judge is NOT consumed here**: its consumer is the N4 (soft-green) rung, so wiring it into
+  a `judged` close stage and executing BA-20's acceptance criteria live (running `calibrate`
+  against the frozen floor and naming the hash it graded) are deferred to N4's opening — a
+  verified delivery, not a verified acceptance.
+
+- **The interview loses its repo question and widens its scope question — hamr drove the terminal
+  himself and ruled on the wording (2026-08-15).** `GREEN_QUESTIONS` drops *"Is there a code repo I
+  can look at? Where?"*: the repository is already MANDATORY STRUCTURED input (`runInterview`'s
+  `repoPath`, taken from `--patient`, without which nothing starts), so asking a person to re-type
+  a path the machine is holding is the SWE tax this product exists to refuse — and it invites a
+  second, drifting answer for one fact (hamr's own live answer was *"Yes — the patient itself."*).
+  Question 2 gains the other half of the same decision — *"…And which files should the work read or
+  draw from (they stay untouched)?"* The sets **renumber contiguously from 1**, so the green set is
+  five questions and hitl's signer-ask is now **6**, not 7. Nothing hardcodes a count: `run-author`,
+  `run-interview` and the suite all read `requiredAnswersFor(cls)`. **Adopter note: an
+  `answers.json` written before this change is SILENTLY MISREAD** — its key `6` was the repo answer
+  and now reads as hitl's signer-ask, and its key `7` is dropped as unasked. Re-run the interview
+  (it is $0 and calls no model) rather than hand-editing.
+- **The interview's own prose stops citing finding numbers and says the mechanics out loud.** The
+  goal prompt no longer prints *"F87"* at a person who has never read the findings — the rule is
+  unchanged and the price is now named as a price. The fence prompt says what a fence IS (writable
+  files, everything else read-only), that patterns are repo-root-relative, and that absolute paths
+  are refused because the run works on a copy. The multi-line hint spells the keystroke: *press
+  Enter on an empty line, i.e. Enter twice*.
+- **Suspended minutes stop charging a signed wall — the run clock is MONOTONIC** (hamr's ruling,
+  2026-08-15: *"if computer suspended, time shouldn't count."*). `createClock`'s default source
+  moves from `Date.now` — CLOCK_REALTIME, which counts every minute the machine was asleep — to
+  `performance.now`, which reads libuv's `uv_hrtime` and so CLOCK_MONOTONIC, which per
+  `clock_gettime(2)` *"does not count time that the system is suspended"*. That was not
+  hypothetical: an s2idle suspend landing mid-run charged **45 phantom minutes against a
+  45-minute wall**, and the same exposure covers an NTP step or a hand-set date. Nothing corrects
+  for a suspend after the fact; the reading simply never includes one. **The epoch change is safe
+  by construction**: every use of `now()` in `src/clock.js` is RELATIVE (`now() - startedAt`) and
+  no value derived from this clock is ever published as a calendar timestamp — `report()` carries
+  durations only, and the `at:`/`ts:` stamps a run writes come from their own `Date.now` seams in
+  `src/reuse.js`, which stay calendar-correct on purpose. `now` is still injectable, which is what
+  every clock test drives. Three runJob-level wall tests mocked `Date` because that is the only
+  seam `runJob` exposes, and the monotonic source made them BLIND — they redded `provider-red` /
+  `plan-red` instead of `wall-halt`, which is the failing-first evidence that the mock was
+  load-bearing; they now go through one shared helper (`tests/helpers.js`) that pins
+  `performance.now` to the mocked `Date` and restores it after. **KNOWN LIMITATION, recorded
+  rather than fixed** (PRD Addendum v1.67, hamr: *"leave it and mark it in prd as possible known
+  limitations"*): the OUTSIDE watchdog still measures silence in CALENDAR time, so a long suspend
+  can still read to it as a stall. The half that charges money is fixed; the half that reads
+  silence is not.
+- **`scripts/run-u.mjs` PRINTS the sleep-inhibitor launch line** the way `scripts/run-reuse.mjs`
+  has since F72, instead of carrying it as a comment — a rule nobody reads at the moment they
+  launch (and like every runner change here, `scripts/` never ships in the tarball). A suspend
+  freezes EVERY guard: the outside watchdog is a POLLER, so it cannot observe — let alone kill — a
+  run whose machine is asleep, and after the monotonic clock above the run's own wall does not
+  count those minutes either. A suspended run is therefore simply UNWATCHED for as long as it
+  sleeps, and the launch banner is the last moment the operator can still do something about it.
+  The printed line is `systemd-inhibit --what=idle:sleep --why="bareloop u run" env …`, with `env`
+  before the key assignment on purpose: `systemd-inhibit` execs a COMMAND, and a bare `VAR=value`
+  prefix is shell syntax it would try to run as one — the key still rides an env assignment and
+  never argv. A pause with no ruling prints the line prefixed to NONE of its three doors, because
+  prefixing one would quietly recommend it and this surface's lean runs the other way.
+- **The hitl terminals get one canonical spelling, and the pause screen gets one renderer**
+  (`/code-review medium`, findings 2 and 3). The terminal names were hand-spelled string
+  literals at every reader: 9 sites in `src/planrun.js`, the ledger's `EXCLUDED_ESCALATIONS`
+  set, `src/reuse.js`, the `src/index.js` re-export and 7 sites in `scripts/run-u.mjs`. They now
+  all import the constants exported from `src/declaredclose.js` (the `hitl-cancel` constant
+  introduced with them is gone again — see the review-door entry above). Beside it,
+  `printPauseEvidence()` in `scripts/run-u.mjs` becomes the SINGLE assembly point for both pause
+  screens — the resume preview and the fresh-pause terminal — so the fallback rule cannot drift
+  between them: with a pause record the package is the close's own ask plus the diff of what
+  changed; without one it says so plainly, rather than rendering the bare "approve?" that ruling
+  2 forbids. Finding 1 — `runReuse` classing a `hitl-pause` as a casualty and buying another try
+  — was parked as arbiter territory when this bullet was first written, and has since been FIXED
+  on this same branch: `d183c95` gave the checkpoint its hard stop (`hardStop` reads
+  `HUMAN_CHECKPOINTS` before the retry table, so the loop hands the run back rather than re-asking
+  a question already put to someone who has not answered), and `4e0ab94` gave the ROW its own
+  `checkpoint` class, so the machinery no longer mis-describes a waiting run as one that died.
+
+### Fixed
+
+- **`--lang` given with no value no longer poisons the command the interview prints.** `arg()`
+  returns the EMPTY STRING for a flag typed with nothing after it and `?? 'js'` passes an empty
+  string straight through, so the printed next-step command collapsed to `--lang  --budget 2.5` —
+  copy-paste it and `run-author` reads the word `--budget` as its language. Both scripts now stop
+  LOUD on a present-but-empty `--lang` (an ABSENT flag still takes the default, the same rule the
+  money ceilings are parsed by).
+- **A crashed authoring run left one line and no body — `scripts/run-author.mjs`'s spine now says
+  it died.** A real run's `author-<runid>.jsonl` held exactly ONE line, `author-start`, and
+  stopped: the paid pipeline threw, the error reached the operator's terminal as an unhandled
+  rejection, and the spine's record of the run was **byte-for-byte what a run still IN FLIGHT
+  looks like**. A log that cannot tell a death from a hang is not a record of either — which is
+  why this is a fix to the record and not a new feature, even though the mechanism is new
+  records. The fallible span — from just after `author-start` through the end of the main flow, a
+  real scout, a real model call and a real toolchain per close stage — is now ONE try/catch, and
+  nothing above it is inside, deliberately: the argv/config `die()` paths run before the spine
+  file exists, and a crash record with no spine to land in is a record nobody can read (those
+  still stop loud on stderr and 2). **The catch RETRIES NOTHING and SWALLOWS NOTHING**: the whole
+  error reaches stderr FIRST and verbatim — before the two writes that could themselves fail,
+  because a diagnosis that only arrives if the disk is writable has a dependency nobody asked for
+  — and only then is it written down as `author-crash` plus `author-end {outcome:'crashed'}`.
+  `crashRecord` lives in `scripts/author-readout.mjs`, where the runner's other
+  untestable-in-place rules live: `name`/`message`/`code` go through `redactSecrets`, and the
+  STACK — the evidence, and the string most likely to quote a credential out of a URL or an env
+  value into a file that outlives the run — goes through `scrubRaw`, the ONE persist boundary,
+  which redacts over the same `SECRET_PATTERNS` inventory the validator reds on and bounds it with
+  the bound announcing its own size (F28). One level of `cause`, no chain-walking. A non-Error
+  throw is recorded honestly rather than coerced into an Error shape. The detail is said ONCE, on
+  the crash record; `author-end` carries only the outcome, because two hand-spelled copies of one
+  fact are two instruments that can disagree (this file already paid for that in the
+  cap-halt/pricing-red type). The spine write is itself guarded — a crash handler that crashes
+  destroys the report it was built to make (F70) — and a failed append says so out loud while the
+  original error still stands. **NEW EXIT CODE 4**, distinct from 1 (a refusal or a failed gate),
+  2 (operator/config) and 3 (a leak): a crash is none of those, and sharing a code with a refusal
+  would file a bug as a result. The leak scan still OVERRIDES a 4 deliberately — a written secret
+  is the harder line, and the crash keeps both of its own louder channels. **No reader changes**:
+  `classifyIncidents` (`src/ledger.js`) keys on event types it does not emit, so both new records
+  fall through every branch and are simply not counted — uncounted, never miscounted, which is
+  the fail-safe direction.
+- **A KILLED authoring run kept losing 100% of its own money record — it now leaves a body, the
+  way a crash does.** The crash catch above covers a THROW and never a SIGNAL, so a `^C` on a run
+  that merely looked hung ended the process with the spine holding one line, `author-start`, and
+  with the spend existing only inside the cost book, which dies with the process (F6/F12: a
+  halted attempt's spend was invisible by 300×). `scripts/run-author.mjs` now handles `SIGINT`,
+  `SIGTERM` and `SIGHUP` by emitting `author-killed {signal, phase, …cost}` plus
+  `author-end {outcome:'killed'}` — `appendFileSync` inside `emit` is synchronous, so the record
+  is on disk before the process goes — then removing its own listener and re-raising the signal,
+  because the honest exit code for a signal death is 128+signo and any code set here would file a
+  killed run under a name this runner's vocabulary already spends on something else. **SIGKILL is
+  not covered and is not pretended otherwise**: it is uncatchable, and there is no handler to
+  write for it. The killed report and the live running total read ONE list (`metered`) through
+  `tallyCalls` — the same reader the library's own cost book uses — rather than a second
+  hand-spelled accumulator, because this file has already paid once for a pair of instruments
+  over one fact (the cap-halt/pricing-red type); `costUsd` rides as `null` when a call was
+  unpriced, so the known half is reported as a `≥` floor instead of `?? 0` laundering unknown
+  into $0. Both reporting seams are wrapped best-effort (`try/catch`) on purpose: progress
+  reporting on a PAID run must never take down the work being paid for (F70), and the run's real
+  records — `authored.json`, the crash catch, `author-end` — are all downstream and unaffected.
+- **`scripts/run-interview.mjs` stops making an offer that can only be refused.** With no
+  `ANTHROPIC_API_KEY` in the shell, `run-author.mjs` exits 2 at its own door before a spine
+  exists, so the *"Run it now? [y/N]"* question had exactly one possible outcome and spent a
+  person's attention on a choice they did not have. Unkeyed, the question is not asked at all and
+  the closing line says which state it is in (*"Not offered — there is no key in this shell to
+  run it with"*) rather than reporting a refusal typed on the operator's behalf. What is printed
+  instead is what is actionable: the exact command, and one line saying to set the key in the
+  shell from their own secret store — deliberately WITHOUT a specific incantation, because which
+  store a person keeps a key in is theirs, and the one thing this script must never do is put a
+  key anywhere a command line can be read from.
+- **A run that landed its own terminal now dates its own stop (N4 surface §1.6).** The resumed
+  leg's `deathAt` preferred the watchdog's kill record whenever one existed. That is right for a
+  run that was KILLED — the process was alive, silently, right up to the signal, and billing only
+  to its last emitted event under-reports the wall it really burnt — and wrong for a run that
+  ENDED ITSELF. A `hitl-pause` is the second kind: a clean terminal a person may answer days
+  later, and any record dated after it bills their deciding time to the run's wall. The POC
+  measured the shape — a report 45 days after the pause put `RESUME_WALL_MS` at 0 and doomed the
+  resumed leg before it opened, on a run that had barely started. So the preference order is not
+  *"the later record wins"* but *"the run's own record wins when there is one"*: a spine carrying
+  a `job-end` returns `null`, which hands `readResume` back its own documented default (the last
+  event's timestamp, which for a clean terminal IS the terminal). The order moved out of
+  `scripts/run-u.mjs` into `deathAtOf` (`scripts/u-readout.mjs`) so a test can reach it at all,
+  with the killed-run case kept as the control, and an unreadable watchdog stamp reads as UNKNOWN
+  rather than handing a `NaN` to a fold (F6).
+- **A pause is a PHASE — the resume readout stops walking off the end of the plan (N4 #9).** hamr
+  read `at step 2 of 1 "(unknown)"` on the first paused resume preview he drove. A pause happens
+  AFTER the plan's steps, at the close's human stage, and the checkpoint a person actually meets
+  reads `phase:'steps'` with every step green (the pause site emits no `outer-close` of its own),
+  so the `at` line's step arithmetic ran past the plan and named an id nobody drafted. The same
+  walk-off was reachable one terminal over with no person involved: a kill between the last
+  `step-end` and the `outer-close` leaves exactly that shape and printed `step 3 of 2`. The
+  rendering moves to `resumeAtLines` (`scripts/u-readout.mjs`, where a test can reach it) and
+  holds one rule — a step count never exceeds the plan, and the phase is said in words: the human
+  review, the close and its fix loop, the close after an exhausted plan, or a named next step.
+  Rendering only; it decides nothing and bounds nothing.
+
 ## [0.10.0] — 2026-08-13
 
 ### Added

package/README.md

@@ -32,20 +32,31 @@ bareloop takes that principle and makes it the product. You describe a repeated,
 job and what *done* means for it. The agent handles the rest, under verification it never
 gets to touch.
 
-## Three kinds of done
+## Two kinds of done
 
-Jobs differ by who or what decides the result is good. That difference is the product's
-spine, so you choose it when you create the job.
+Jobs differ by what decides the result is good. That difference is the product's spine, so you
+choose it when you create the job.
 
 **green** — a predicate decides. Tests pass, the build exits clean, the type checker is
 quiet.
 > *"Fix the failing suite."* · *"Migrate this package to strict types."*
 
-**softgreen** — an assessment decides against a rubric.
-> *"Reserve a ticket from SF to LA under $400."*
+**softgreen** — an assessment decides, against a rubric you wrote and signed. You say what
+separates a pass from a fail, and give one example of each; the judge only quotes back what it
+finds in the real work, and fixed rules — not the judge — render the verdict.
+> *"Reserve a ticket from SF to LA under $400."* · *"Improve my resume."*
 
-**hitl** — a human makes the final call.
-> *"Improve my resume."*
+## You have the last word
+
+Every judged run ends at a **review door** (a purely mechanical green finishes on its own —
+ask for the door with `--review-door`): here is what changed, here is what every check said,
+here is what I'm asking you. **Accept** it, **rerun** it — what you type is the feedback the next
+attempt works from, funded fresh — or **pause** it and decide another day. A pause costs
+nothing, keeps everything, and picks up from the start of the last step whenever you come back.
+Never come back, and it quietly expires on its own.
+
+The door never rewrites the run's own verdict. A green stays green in the record whatever you
+decide about it; your answer says what happens to the work, not what the work scored.
 
 ## Why the agent authors the workflow