@techgoblin/gobstack 0.0.0-stage → 0.4.4-beta.2

package/docs/LIMITS.md

@@ -0,0 +1,591 @@
+# Limits — where this is weaker, and what is unproven
+
+Honest accounting. Nothing here is a hedge for a defect that could be fixed; each is either a
+deliberate trade or an unfilled gap.
+
+## Weaker than the Cursor harness it learns from
+
+1. **No live-drive lane is shipped.** A harness that launches and drives the real application is
+   what several of the source harness's playbooks depend on, and it is inert without a plugin
+   that is not shipped. `P6` can *author* a project-local driver; goblin-stack cannot ship one,
+   because that is per-project work. **The "live lane is the floor" rule is stated and
+   unfilled.**
+2. **No cloud agents.** There is no per-agent computer here, so nothing that needs a machine of
+   its own.
+3. **No agent graph and no verdict-ledger daemon.** The board replaces both at a lower
+   resolution: `parents` expresses ordering, not data flow, and sibling cards cannot see each
+   other.
+4. **Fifteen playbooks against twenty-three.** The cuts in `docs/FLOWS.md` are deliberate and each
+   is argued, but real coverage is lost: performance hillclimbing, pixel parity, trace
+   forensics, stack landing, worktree hygiene.
+5. **No swarm or arena fan-out.** The read-versus-write axis says that is correct for this work
+   mix; it is still a capability the other harness has and this does not.
+
+## Weaker than the status quo it is meant to improve
+
+6. **The biggest measured defect is out of reach.** The orchestrator's routing text says a bare
+   subagent spawn reaches the specialist profiles; it does not. That is the highest-cost defect
+   on the box, and goblin-stack cannot fix it from inside a project repo. It ships corrected
+   text in `docs/INTEGRATION.md` and lints its own artifacts (`MD-03`); the fleet-side edit is
+   escalated.
+7. **The staleness of a fleet-config repo is detectable and not fixable here** — the E-class gate
+   notices it, and the underlying job bug belongs to another repository.
+8. **A new surface to maintain.** Per-repo vendored `.goblin/` plus `.hermes/skills/` means
+   upgrade debt in every adopted repo, plus one more command pair to learn. The counter is that
+   the alternative — profile copies — already failed.
+9. **Ten rows are labelled `advisory`, and nine of them are prose with no check at all:**
+   `HP-04`, `HS-03`, `CM-02`, `MD-03`, `PG-04`, `DOC-01`, `DOC-02`, `SC-09` and `JG-03` (G2, the
+   judge lane that has never returned a non-`done` verdict). One (`MD-02`) is
+   advisory-labelled but still reports its state as `ADV`. Each is counted and capped, but a
+   counted rule is still not an enforced one, and **the cap is a policy, not a proof.**
+10. **It does not reduce the profile-skill surface.** It only stops that surface from growing
+    with project procedures; the existing divergence is a separate cleanup.
+11. **`HS-02` cannot prove anything until a round lands.** With no pinned pre-change commit it is
+    skipped with a reason. That is honest, and it means a brand-new install has **no** REPLAY
+    evidence at all. The shipped `checks/assert.mjs` is a scaffold: it asserts something true by
+    construction, so `HS-02` correctly reports it as a check that proves nothing until it is
+    replaced with real probes.
+12. **`HP-03` proves a date exists, not that a number is fresh.** A HANDOFF can carry yesterday's
+    number with today's date and pass. Since V1 the row is anchored on the gate names
+    `.goblin/goblin.yaml` DECLARES and skips the template's own example sentence, so a real gate
+    line can no longer lose its `measured <date>` in silence (G8-2 measured the old row passing
+    exactly that); but nothing re-measures the number, and a gate-bearing line that names no
+    declared gate and carries no gate-shaped keyword is still unseen.
+13. **`PG-04` is documentation, and `PG-05` is a text reading — re-declared at W4 rather than
+    called a gate.** `PG-04` stays `advisory` because the forge is unobservable from inside a
+    repo: a protected branch whose only admin is the person pushing protects nothing, so a
+    required check armed under the sole admin's own identity binds nobody — the second half of
+    that sentence is measured (Harvey is the sole admin of every repo he owns), and changing it
+    is a forge/account change, not a repository change. `PG-05` **used to be a heuristic**: it
+    counted "every step is guarded", so a single **job-level** `if:`, a job with no step, and the
+    measured real shape (one *unguarded* step deciding whether the guarded gate step runs) all
+    passed it while GitHub reported Success. It now refuses a conditional job, a conditional
+    step, a job that declares no step and a workflow with no `jobs:` — see `docs/CI.md` §1 for
+    why a skipped job is a green light. It is still a **text reading**: no YAML parser
+    (`docs/CONTRACTS.md` allows none), so a flow-style `jobs: {…}` mapping is refused rather than
+    parsed, a `#` inside a quoted string is read as a comment, and a conditional step that is
+    genuinely safe is indistinguishable from the trap. It cannot see branch protection, the
+    required-check list, or whether the job ever ran — `PG-04` is the row that says so. `PG-06`
+    is the other half and has its own limit: it proves the declared gate is **invoked** in a file
+    under `.github/workflows/`, never that the forge marks that job required, never that it is
+    the job the forge waits on, and never that the workflow can fail.
+14. **`CM-02` cannot be enforced.** A backtick lost to command substitution leaves no trace a
+    later check can read.
+
+## Unproven at the level that matters
+
+15. **The controlled evidence now exists, and its counter-finding is about this repo (corrected
+    2026-09-25).** This entry used to say that **no controlled study of skill efficacy exists**, so
+    that nothing here claimed the installed skills change agent behaviour. That is out of date:
+    **SkillsBench 1.1** (benchmark published 2026-06-16) reports that **curated Skills raise the
+    mean task-resolution rate from 33.9% to 50.5% — +16.6 points, a 25.5% normalized gain — across
+    87 tasks, 8 domains and 18 model–harness configurations**, with every one of the 18
+    configurations higher with Skills (configuration-level gains from +4.1 to +25.7 points).
+    **The counter-finding matters more here than the headline:** in the same benchmark's
+    **self-generated condition — the agent authors its own Skills before solving — all three tested
+    configurations landed BELOW their no-Skills baseline** (−8.1, −11.3 and −11.5 points), while
+    curated Skills stayed above it. **goblin-stack installs agent-authored skills**, so the lower
+    row of that result is a warning about its own output, not someone else's: a generated skill
+    accepted after a skim is a different proposition from a written one. It is why `P6` hands the
+    generated skill to `P12` before any `verified:` date advances (`docs/RISKS.md` K15).
+    **`P12` is still the mechanism to find out, and it has still never been run**: the record format
+    is now specified (`skills/goblin-eval/SKILL.md`) and **no row reads a lane** (#31).
+    Pin: `https://www.skillsbench.ai/blogs/skillsbench-1-1`, sha256 of the retrieved page
+    `d812bb7c2da702cc67556eb5f46b9e93faaca19d3545376544866e940126ef15`, fetched 2026-09-25; the
+    eleven-token ban and the judge procedure are argued in
+    `https://ai.engineer/talks/0vphxNt4wyk-don-t-ship-skills-without-evals`. **Not registered**:
+    the source registry is `manifest/sources.tsv` (G9's artifact) and that file does not exist in
+    this repo — measured, `git ls-files manifest/` names five files and none of them is
+    `sources.tsv` — so the registry row is composed in the W2 report for G9's patch instead of
+    being written into a file this card does not own.
+16. **Every number in the design is a file read or a third-party published number; none of it is
+    a measurement of this harness under load.** The harness was built and its verifier was shown
+    RED under a deliberate break; that is a statement about the mechanism, not about outcomes.
+17. **One spec deviation, recorded rather than hidden.** The design spec's literal HANDOFF check
+    is `grep -q "$(git rev-parse --short HEAD)" HANDOFF.md`, which **can never pass** — committing
+    the HANDOFF moves HEAD, so the file can only ever name an ancestor. `HP-05` is therefore
+    mechanised as *the HANDOFF names a commit that exists in this repo and is an ancestor of
+    HEAD*, which still catches the defect the rule exists for (an artifact that names no commit
+    at all). The deviation and its reason are in the row's own `if_not_why` column.
+18. **`.goblin/installed.json` is not signed, so nothing here proves it was not rewritten.** Every
+    drift check — `IN-02`, `SK-02`, and `HS-01`'s hash of the harness dir — reads its expected
+    hash out of that one file, and that file is the one file no check protects. Measured: append a
+    byte to `.goblin/bin/goblin-verify`, rewrite its recorded hash in `installed.json`, commit, and
+    the run is **fully GREEN** (`43 passed, 0 failed`, exit 0). One edit defeats three rows at
+    once, and it is the cheapest way to fake a green run. Doing better needs an anchor the target
+    cannot edit — a signature, or a hash held outside the repo — and goblin-stack has no such
+    trust root: the source checkout is not guaranteed to exist at verify time, and any value
+    stored in the tree is editable by the same hand. It is therefore **recorded here and printed
+    in the "cannot see" footer on every run**, not claimed away.
+19. **No automation has ever run here.** Cost per run, the respawn guards under a nightly
+    producer, and whether `researcher` is the right reporter profile are all unmeasured; the
+    first watched run of A-02 is what produces those numbers. The producer's own ceiling is a
+    **run count**, not a dollar figure — no config key holds the spend cap, so no automation can
+    read it, and none pretends to.
+20. **The dedup key is a dedup, not a mutex.** The board's lookup runs before the write
+    transaction, so a concurrent create can insert twice and the next lookup stabilises on the
+    newest. The key stops a duplicate storm; it does not make one impossible.
+21. **`AU-02`'s normalisation is only measured on synthetic reports.** Two differently-typed
+    copies of one symptom give one key and a different symptom gives another, but whether a real
+    report set normalises well enough is unknown. Its failure mode is a duplicate card, never a
+    lost report.
+22. **No audit has ever run against a real registry here.** Everything `SC-07` does was measured
+    against a canned npm-audit report (`tests/t-audit.sh`), so the parse is proven, the *policy*
+    is not: whether the recorded waiver set matches the real advisory set is unknown until the
+    first real `goblin-audit`. Its skip-with-a-reason behaviour on a repo with no record is what
+    keeps that honest rather than silent.
+23. **The audit record is read by field name, not by a JSON parser.** `goblin-audit` recognises
+    npm's `vulnerabilities` / `via` shape (`source`, `name`, `url`, `range`) with awk and REFUSES
+    (exit 5) anything it cannot parse, rather than writing an empty record that `SC-07` would read
+    as clean. A different audit tool is therefore a refusal, not a silent pass.
+24. **The freshness rows need a GNU `date -d`.** `SC-07` parses the record's date that way; on a
+    host without it the row FAILS with the reason rather than assuming the record is fresh.
+25. **`SC-04` reads one statement, not one program.** A cookie write spread over three lines (or
+    assembled through a helper) is not seen, and the row says so in its own cell.
+26. **The last advisory slot is an open decision, not a rule.** Measured (V1): the advisory rows
+    are `HP-04`, `HS-03`, `CM-02`, `MD-02`, `MD-03`, `PG-04`, `DOC-01`, `DOC-02`, `SC-09` — 9 at a
+    ceiling of 10 — so **exactly one slot was free**, and `SK-03` reported that arithmetic at the
+    time (`advisory 9 of ceiling 10 (1 free slot)`; **corrected 2026-09-25 (AB3):** the run prints
+    `advisory 10 of ceiling 10 (0 free slots: the next advisory row FAILs)`, W3's `JG-03` having
+    taken the slot — the two notes below carry the chronology). Two planned cards each wanted the
+    slot: G1's `FM-03` (the feature map) and G2's `JG-03` (the judge agent). **Nothing in this repo
+    chooses between them**, and V1 deliberately spent nothing. The cap is a count, not a strict
+    bound: measured, 10 advisory rows at a ceiling of 10 **pass**, and the 11th FAILs (10 at a
+    ceiling of 9 FAILs). So the tenth row is allowed; the eleventh is not. Whoever lands second
+    brings a real command.
+    **Decided 2026-09-25 (W2):** G1's `FM-03` does **not** take the slot. The feature map ships
+    `FM-01` and `FM-02` as real commands, and the one thing they cannot check — whether the map
+    lists every feature — is recorded as #30 instead of as a counted row; the slot is left free for
+    G2's `JG-03`, whose judge is the mechanism `P12` actually needs. Measured after W2, `SK-03`
+    still reads `advisory 9 of ceiling 10 (1 free slot)`.
+    **Spent 2026-09-25 (W3):** G2 landed and `JG-03` took it. The slot is now **full** — measured
+    `advisory 10 of ceiling 10 (0 free slots: the next advisory row FAILs)` — so the next author who
+    wants an advisory row must raise `advisory_ceiling` in the same change and write down why,
+    rather than discovering the cap from a red run. This is not a rule change; it is the arithmetic
+    the ceiling was always meant to force into the open.
+30. **The feature map is an inventory, and nothing checks that it is complete.** `FM-01` checks the
+    index against the feature files that exist and the four-H2 entry contract; `FM-02` is a tripwire
+    over `entry_paths:`. **Neither can see a feature nobody wrote down** — that needs semantic
+    judgement over the app, which no command here performs. Three smaller blind spots are named in
+    the rows' own why-cells and repeated here: `FM-02` searches for the token under `source_root:`
+    and therefore **reads a vendored copy, a lockfile or a build artifact as "still resolves"**; it
+    sees a *file* change and not a *behaviour* change, so a refactor that leaves the route alone
+    reports stale-and-wrong and a behaviour change in a file the token does not appear in is missed;
+    and a `verified:` date is a **claim the row cannot test** — nothing distinguishes a feature
+    driven that day from a date typed that day. The first of those is why `source_root:` should name
+    the source tree, not the repo root, and the third is why the upkeep pass in
+    `skills/goblin-feature-map/SKILL.md` requires the date to advance only for a feature someone
+    actually drove.
+31. **The P6↔P12 loop is wired as a contract with no runner.** `P6` now hands a generated
+    verification skill to `P12` and `verified:` does not advance until an eval record exists; the
+    record's shape, the eleven-token ban, the cheap-checks-first ladder, the merge rule and the pass
+    condition (every seeded defect detected, the control's number at zero, every correction RED
+    before GREEN) are all specified in `skills/goblin-eval/SKILL.md`. **No row reads a lane, and
+    nothing executes an eval** — measured after W2, `manifest/enforcement.tsv` has no `EV-*` row,
+    and `P12` has still never been run. The runner and the record checks (`EV-01`..`EV-04` in G1's
+    design) are deferred to a follow-up card, deliberately and in the open, rather than half-built:
+    a row that reads a record nobody writes would pass vacuously and look like enforcement.
+32. **The judge is a language model grading prose, and the record proves a handle exists — never
+    that the handle supports the verdict.** `JG-01` FAILs a `done` verdict whose evidence resolves
+    to nothing: `sha:<hex>` must be a commit in `git rev-list --all`, `file:<path>` a path under the
+    root, `sha256:<hex>` the digest of a file under `.goblin/loop/`. All three say *exists*. A judge
+    may cite a real commit that has nothing to do with the claim and pass, and a `cmd:<command>`
+    token resolves **nothing on purpose** — the command ran, its output is not in the record, and a
+    verdict resting on it is the self-report the row refuses. Two smaller blind spots are stated in
+    the row's own why-cell and repeated here: the row cannot see **which lane returned a verdict**
+    (no file in a repo observes the profile that ran — `MD-03`), and **an unresolved judge lane is
+    an `ADV`, not a failure**, because a repo cannot choose the fleet's routing (`docs/ROLES.md`,
+    "the measured caveat"): measured on this box, the judge lane resolves to no profile at all.
+    Finally, the `JG-03` counter-measure is **policy, not a check** — one known-red control verdict
+    per wave, recorded in `docs/LOOP.md`: a lane that has judged twice cannot be called always-yes,
+    and the history that would show a bad lane lives across cards and repos.
+33. **`LP-04` measures a changed evidence pointer, which is a proxy for progress — not progress.**
+    Three consecutive verdict rows with an identical non-empty pointer and a result that is not
+    `predicate:green` is a FAIL naming the row numbers, and that is the strongest thing a repo can
+    read without running the loop. A loop that edits a file each turn to keep the pointer moving is
+    not caught, and **nothing in Hermes detects a lack of progress either**: measured in
+    `hermes_cli/goals.py`, `run_kanban_goal_loop` carries no progress state at all — its whole
+    state is `last_response`, `turns_used` and `nudged_to_finalize`, so a loop that returns
+    `continue` for the same reason nineteen times spends nineteen turns and then blocks
+    (`hermes_cli/goals.py:1636-1638`, `:1689-1696`). So the budget is the backstop, and the budget
+    has its own blind spots: `LP-03` proves the declared budget is a positive integer at or under
+    `loop_max_turns_ceiling` and that the record holds no more verdict rows than the budget — it
+    cannot see whether the budget is **affordable**, and cost is not a field the record holds
+    (neither turns nor tokens nor the per-turn auxiliary judge call). Three further "not yets" are
+    structural rather than measurable: `LP-01` proves a recorded first run exists and that its
+    timestamp is at or before the first log row — **not that the command ran and not that the
+    `exit=` value was measured rather than typed** (`HP-03`'s defect, one artifact over); `LP-02`
+    proves the predicate file still hashes to its pin — **not that the predicate is the right one,
+    and not who edited it**; and a predicate that was **vacuously true from the start** (a `grep -c`
+    against a renamed directory) passes `LP-01` and `LP-02` and ends the loop green on nothing.
+    `LP-05` makes a write-up mandatory and therefore visible — nothing can make it true.
+
+## What the harness refuses to do
+
+It does not claim a green run means the work is right. `goblin-verify` asserts that the installed
+files are the files on disk, that every rule with a command still passes, and that the
+untestable remainder is counted and capped — and it prints, on every single run, what it cannot
+see: the five upstream blind spots, plus the ban lane's own (the unsigned ban table #28, the
+text-probe gap #27, and a ban that is invisible until verify runs, `V3-1`), plus the judge/loop
+lane's (#32: a handle that exists is not a handle that supports the verdict; #33: a changed
+pointer is a proxy for progress, not progress), plus the CI lane's (#34: a file is not a gate —
+the required-check list, the bypass switch and the push identity are forge state; #35: the
+Electron perf number is a host gate, and the ratchet deliberately carries a different metric).
+
+27. **The ban probes are text probes, not ASTs.** `BN-01`, `BN-02`, `BN-03` and `BN-05` are
+    `grep` over source under `bash`/`grep`/`awk` only — the dependency contract in
+    `docs/CONTRACTS.md` allows no parser and no `npm`. So a `: any` inside a string or a comment
+    is reported, `Record<string, any>` (no leading colon) is missed, and BN-05 does not resolve
+    module aliases or dynamic imports. The AST-grade form of the same bans (BN-01..BN-04 in
+    `G5.md` §C.2) needs ESLint and `dependency-cruiser`; that is why **G5's `BN-04` (the nine
+    named unnecessary-effect patterns) is NOT shipped** — it cannot be mechanised without a
+    parser, and a ban that cannot go red is worse than advisory, so it is recorded here rather
+    than as a row that would cost the last advisory slot. A text probe with a stated
+    false-positive set is still a gate: each BN row is shown RED under its own violation and
+    GREEN when it is removed (`tests/t-verify-red.sh`).
+28. **The ban table's integrity rides on `IN-02`, and `IN-02` rides on an unsigned record.** A
+    project could empty `manifest/bans.tsv` (or edit a `detect` command) and the ban gate would
+    pass vacuously — `BN-00` fails closed on an *empty* table, but it cannot see a table whose
+    rows were weakened, because `.goblin/manifest/bans.tsv` is hashed by `IN-02` and
+    `.goblin/installed.json` is not signed (`docs/LIMITS.md` #18). The ban list is not
+    tamper-proof; it is as strong as the record every drift check trusts. **Measured (W5-12):** with
+    `BN-01`'s `detect` cell set to `true`, `BN-00` and `BN-01` both PASS and the only row that
+    changes verdict is `IN-02`'s drift check — the `W5-12` control in `tests/t-verify-red.sh` pins
+    exactly that, and nothing else can. Z1's verdict is to **record this, not fix it**: the `detect`
+    cell is executable content, and a row able to judge whether another row's command *means*
+    something would be an `eval` over the matrix, which the same card rules out. The gap is now
+    stated, measured and controlled one row over, which is the most this table can do without
+    becoming an interpreter.
+29. **The perf ceiling must EQUAL the recorded baseline, so a budget with headroom is not
+    expressible.** `PF-01` FAILs when `ratchet.ceiling` and `perf.baseline_value` disagree
+    (G8-6b), which is what stops a one-line ceiling raise from passing while printing the
+    contradiction. The cost is real and the row cannot see it: a project that wants the ratchet
+    to allow, say, 10% growth over the measured baseline cannot write `ceiling: 44000` beside
+    `baseline_value: 40000` — the row reads that as a disagreement. Headroom is expressed by
+    re-anchoring BOTH (measure on a pinned commit, then set `baseline_value` and `ceiling` to the
+    new number together, and record it as an operator action), which is the deliberate re-anchor
+    the row's own why-cell names. It is a bound on the budget's shape, not a claim that the
+    budget is the right one — and it still never re-measures.
+34. **The CI lane reads files, and a file is not a gate.** The shipped workflow
+    (`templates/ci/goblin-gate.yml.tmpl` → `.github/workflows/goblin-gate.yml`) has no `if:` at any
+    level, but nothing in a repository can make GitHub **require** it: the required-check list,
+    the bypass switch and the push identity are forge state (four settings, `docs/CI.md` §1).
+    `PG-06` proves the whole declared gate set is invoked in a file under `.github/workflows/` and
+    cannot prove that file is the one the forge waits on; `PG-05`'s reader has no parser, so a
+    flow-style `jobs: {…}` mapping is refused, and a `#` inside a quoted string truncates the line
+    it is on. Two further measured gaps: the template's job is `ubuntu-latest` with no cache, so a
+    repo whose gate needs a display, a licence, a GPU or a signed-in session cannot use it at all
+    (that is a **host** gate — the class-C rule, restated for class F), and a private repo's
+    Actions minutes are billed to the account (2,000/month free). Measured ground truth at W4:
+    **one** first-party workflow exists in the whole estate and it self-skips; five of the six
+    repos with a remote have none.
+35. **The Electron perf number is a host gate, and the ratchet carries a different metric.**
+    `presets/F-electron.yaml` declares `main_thread_busy_pct` as `perf_host_gate:` and uses
+    `app_bundle_bytes` for `ratchet:` — a **deliberate deviation** from G6 §B.3, which put the FPS
+    number in the ratchet. The instrument that produces it (CDP `Performance.getMetrics`, or
+    `app.getAppMetrics()[i].cpu.percentCPUUsage` inside a real Electron) needs Playwright or
+    Electron plus a GUI, and a shipped rule may use nothing but bash/git/awk/sed/grep/python3
+    (`docs/CONTRACTS.md`) — so `ratchet.cmd` pointing at the probe would make a fresh install
+    **born RED**, which is the one thing the install path must not produce. The probe belongs to
+    the project; a project whose CI needs npm runs it in **its own** workflow. What the number
+    means has a limit of its own, and it is measured: this box is an LXC with no display and no
+    system Chromium, so the G6 sweep came from a *bundled headless* Chromium with no compositor
+    and no vsync. Relative comparisons on one machine are meaningful (which is why
+    `main_thread_busy_pct` works at all, and why frame time does not — p50 stayed flat at 16.70 ms
+    while the main thread went from 1.8 % to 54.5 % busy, `docs/CI.md` §3.3); an absolute FPS
+    claim is not. Three further Electron failure modes are **recorded, not mechanised**, and
+    `docs/CI.md` §4 says why: the dependency-graph boundary check, `ipcMain` sender validation,
+    and fuses at package time.
+36. **A ban's exemption reaches the probe through its environment, so a custom probe can ignore it.**
+    `bans_exempt:` and the inline `// BAN-OK(<id>): <reason>` are filtered *before* the exit code is
+    chosen, because a filter applied to a probe's stdout afterwards cannot change a verdict — that
+    was W5-1, and the old code made every exemption a permanent RED. The engine therefore exports
+    `GOBLIN_BANS_ID` and `GOBLIN_BANS_EXEMPT` and the two shipped probes honour them. A project's
+    **own** `detect` command that ignores the variables keeps the old behaviour: a violation inside
+    an exempted path stays RED. That direction is **closed**, never open, and it is the trade this
+    choice makes. The engine's own stdout is no longer filtered at all, so a probe that ignores the
+    contract prints the exempted lines it reported — loud, and still a FAIL.
+37. **`FM-02` refuses the harness, not every non-source file.** The search skips `.git/`,
+    `.goblin/`, `.hermes/`, the declared `harness_dir` and the map's own directory, so a stub map
+    whose token occurs only in the install no longer resolves (W5-4: `entry_paths: [export]` used to
+    "resolve" to `./.goblin/bin/goblin-verify`). It does **not** exclude the target's own `docs/`,
+    `tests/` or build output: a token that occurs only there still reads as resolved, and excluding
+    them would be a guess about a layout goblin-stack does not know. Nor is there a frequency bound
+    — a common token ("export", "main") is satisfied by the first of hundreds of files.
+38. **The judge lane's model family is reported, never enforced — and on this box it is not even
+    mapped yet.** `JG-02` proves the declared profile *names* are disjoint; W5-6 measured that a
+    profile mapped to the author's own model passed it. `MD-02` now resolves the judge lane and
+    compares its model with the code lane's, so the state is loud — but it stays `advisory`
+    (`return 2`, never a FAIL): goblin-stack cannot choose the fleet's models, and a repo-local file
+    cannot observe *which* model a lane actually ran. Measured on this box at X1: Harvey's
+    `fleet-model.yaml` names **no `judge:` profile at all**, so `MD-02` reports the judge lane
+    **unresolved** and prints the one-line remedy, while every profile it *does* name — nine of
+    them, the named per-fleet role profiles this setup routes its lanes through —
+    resolves to one model under an active promotion. So on this box the honest reading is: the judge
+    lane is not mapped, and the moment it is mapped it will be the author's own family. The
+    comparison is exact model equality, not a version-stripped "family": two spellings of the same
+    family that differ only in a suffix would read as different.
+39. **`LP-02` cannot tell a weaker predicate from a re-scope. The close-and-reopen is recorded, not
+    prevented.** A loop could archive its bar under `closed-<date>/`, write a weaker one, re-pin,
+    and pass `LP-02` and `LP-05` with nothing in the record (W5-7, measured). The row now requires
+    every archive to hold its predicate **and** the pin it was closed under, and the live
+    `predicate.sha256` to name the archived digest on a `previous:` line — so a **silent** relaxation
+    is caught and a legitimate re-scope costs one line. Whether the new predicate is weaker, and who
+    edited it, is not decidable from a digest, and a predicate that calls a script elsewhere is
+    pinned only at its call site. The residual is the same one `LP-01` carries one file over: the
+    record is evidence, not proof that the loop stopped for the right reason.
+40. **A fresh install into a repo with no commits is born RED, and X1 did not change that.** Measured
+    at X1: `goblin-install --class A` into a `git init` with zero commits, then `goblin-verify`, gives
+    `37 passed, 6 failed, 11 advisory, 24 skipped`, exit 1 — `HP-05`, `SP-02`, `GT-02`, `CM-01`,
+    `CM-03` and `PT-02` all read a HEAD that does not exist yet. The install never creates the seed
+    commit (`bin/goblin-install` writes files and stops), and it must not: a tool that commits into
+    Harvey's repo on first contact is the overreach `docs/CONTRACTS.md` rules out. Carried from W5
+    §5's `G8-9` rather than fixed here — it is a **sequencing** limit, not a hole in a row: one commit
+    clears all six, and `tests/t-verify-green.sh` seeds one before it installs.
+41. **Eleven of Y1 §7's eighteen documented mechanisms still carry no control of their own, and one
+    exemption is by design.** The census, stated separably so a reader can check the arithmetic:
+    **18 listed · 2 fixed · 2 closed · 3 controlled · 11 recorded.** Y1 §7 listed eighteen
+    mechanisms; Z1 **fixed** the two load-bearing ones (17, the unrendered token — Z1-3; 18,
+    `replay.cmd` — Z1-4, which now have controls and therefore sit *outside* the "carry no control"
+    set) and **closed** two more (4, the ban engine's `exit 2` paths; 5, `--list` — three assertions
+    in `tests/t-verify-red.sh`, RED against a deliberately broken copy of `bin/goblin-bans`, which
+    is PR-03's second branch: those mechanisms *worked*, they were just unguarded); AA1
+    **controlled** three (1, 2 and 6 — the ban-engine cluster, below). The remaining **eleven** are
+    **recorded here rather than mechanised**. Each was measured WORKING, so an assertion would pin
+    behaviour that already holds, and this box does not have eleven controls' worth of budget; the
+    cost of the gap is exactly the shape of both regressions in this repo's history — a mechanism
+    documented, and nothing asserting it.
+      - items 1, 2, 6 — `bans_exempt:` on the layer probe, the engine→probe
+        `GOBLIN_BANS_ID`/`GOBLIN_BANS_EXEMPT` contract, and its segment alignment: **CONTROLLED at
+        0.4.2 (AA1)** — seven controls in `tests/t-verify-red.sh`, in both directions each (the
+        layer probe's exempt path PASSes and the same crossing import outside it FAILs; a prefix
+        covers its own subtree but does not swallow `src/okay`; a prefix written with a trailing
+        slash exempts nothing; and a project's own probe that exits 0 only when both environment
+        variables arrive). This was the cluster Z1 named as the one with a real engine underneath,
+        and `grep -c GOBLIN_BANS_EXEMPT tests/` was **0** before it.
+      - item 3 — a project's own probe that ignores those variables fails closed: the mechanism is
+        this file's #36, where the FAIL is measured.
+      - item 7 — `BAN-OK` must sit on the offending line: the marker's *effect* is controlled, the
+        on-this-line clause is not (a marker on another line does not clear the violation).
+      - items 8, 9 — `SC-05`'s `.goblin/boundary-waivers` and `SC-08`'s
+        `.goblin/install-hooks.allowlist`: both files are copied and both FAIL directions are
+        controlled; neither *honoured* direction is. (`SC-08`'s reader is the row the minified
+        lockfile defeated — Z2-2; that was a shape, not this item, and it is fixed and controlled.)
+      - item 10 — `SC-03` clause 2, `sec_build_output`: a literal in `dist/assets` FAILs; only the
+        clause-1 path is controlled.
+      - items 11, 12, 15 — `scaffold_checks:`'s SKIP branch, `perf_host_gate` (#35) and
+        `templates/loop/*.tmpl`: declared no-ops (SKIP, a host gate, and templates `docs/LOOP.md`
+        says nothing installs). A control here would assert that nothing happens.
+      - item 13 — `docs/CI.md`'s flow-style `jobs: {…}` refusal: fails closed already, so a control
+        would pin the strict direction of a check that cannot pass vacantly.
+      - item 14 — `bin/goblin-model`: works (`code` → the resolved lane, unknown role → exit 2);
+        only its *absence* from an install is asserted.
+      - item 16 — `docs/LOOP.md` §6's "one known-red control verdict per wave, recorded in this
+        file": a prose obligation, not a command. It is honoured in the write-ups (or not) and no
+        check can read a wave.
+    **W5-10 is answered here too.** The tenant strings `PT-01` forbids do not reach `docs/`, because
+    `docs/` is never installed: the row's directory list is `skills manifest bin templates presets
+    .goblin .hermes`, so a string under `docs/` is source-tree prose that no operator's repo ever
+    receives. That is the whole reason, it is deliberate, and the row is not weakened by it — Y1
+    agreed, and Z1 leaves it. This is the sentence Z1 added so the reason is visible in the shipped
+    artifact rather than only in the wave's own notes.
+
+42. **The doc walk cannot read a path broken at the directory/name boundary.**
+    `tests/t-doc-promises.sh` reads a command directory only as `.goblin/bin/` or `bin/` **with its
+    trailing slash**; when a line ends on the bare directory (`.goblin/bin`, `bin`) and the slash
+    leads the next line (`/goblin-doctor`) or is dropped (`goblin-doctor`), the directory and its
+    continuation are both invisible — the first fragment stops before the name the grammar needs, the
+    second is not preceded by `bin/`, so neither is a token and neither is asserted. Measured
+    2026-09-26 (AB7): a plant of that form in `docs/CI.md` is reported **`PASS`, rc 0**, mentioning
+    the plant **zero** times, and a line that ends on the bare directory with nothing after it is
+    silent too, because the `dangling()` guard needs the trailing slash as well. **0 live instances**
+    — measured: `grep -rnE '\.goblin/bin$|bin/$' $(git ls-files)` → **no output, exit 1**. **What it
+    costs:** a false path written in that form — the exact defect this file's whole species is named
+    for — passes the watch silently, so a future document could hand a reader a command that does not
+    exist and nothing in the walk would say so. **Ticketed, not gated:** reading the form is a change
+    to the tokeniser inside `tests/t-doc-promises.sh`, not to a document, and it is a separate card;
+    this entry records the boundary so a green run cannot imply a coverage it does not have. The
+    control's own header now names the form as uncovered instead of claiming a sensitivity it does
+    not have, and the slash's load-bearing role is stated where the dangling rule is.
+
+43. **The engine footer names its judge, and the statement is unsigned — one forgery covers N
+    repos.** W1's engine split prints `cli_sha256` and `enforcement_tsv_sha256` on every run
+    (and `installed.json` gains an optional `engine:` block recording the same pair), so a repo
+    can state WHICH engine judged it. Nothing signs either hash: the footer is printf output of
+    the very binary it names, and the `engine:` block is a JSON stanza inside the same unsigned
+    record `docs/LIMITS.md` #18 already covers. An edited engine, or a hand-written block
+    claiming a mode that was never resolved, prints whatever it likes — and because the
+    statement now rides in every repo's run, one forgery propagates to every repo that trusts
+    it, which is strictly worse than #18's per-repo record. The old defences still hold and are
+    what the footer must not be read to replace: `--source` pins the engine by flag, a vendored
+    engine keeps beating the global one in the resolution chain, and IN-02 still hashes the
+    repo-local bytes. What is NOT fixed (G6, the no-signing non-goal): no trust root, no
+    detached signature, no third-party attestation. Measured with W1: tamper one byte of the
+    vendored manifest and the footer's `enforcement_tsv_sha256` moves — the footer is a real
+    fingerprint of what ran, it is just not PROOF of it.
+
+## Verdicts recorded, not built (the note-8 questions)
+
+Two library questions were investigated to a verdict and **deliberately built nothing here**, so a
+later session does not re-derive them. The measurements and sources are in
+`goblin-stack-research/G6.md` Part C; this is the durable half.
+
+**Pretext (`chenglou/pretext`) — USE, narrowly, and not as a runtime dependency.** Verified from
+the source of truth: a pure JS/TS library for multiline text measurement and layout, MIT, that
+*"side-steps the need for DOM measurements (e.g. `getBoundingClientRect`, `offsetHeight`), which
+trigger layout reflow"*, using the browser's own font engine as ground truth. Author confirmed
+(Cheng Lou, `_npmUser: chenglou`); the README credits Sebastian Markbage's earlier `text-layout`,
+whose repository now says *"This project is archived. The ideas here evolved into Pretext"*.
+Two verified corrections to the shipped skill: the skill pins `@chenglou/pretext@0.0.6` while npm's
+latest is **`0.0.9`**, and *"15KB zero-dependency"* describes the **runtime**, not the install
+(published tarball `unpackedSize: 887142` across 69 files, `sideEffects: false`, subpaths `.` and
+`./rich-inline`). The concrete use is a **dev-time / harness-time label-fit check** for
+`diagram-studio` and `goblin-ui` — "does this string fit this box at this font", without a layout
+read, as a `checks/*.mjs` assertion with a REPLAY — in exactly the repos that today hand-roll that
+arithmetic (26 `getBoundingClientRect()` calls, no `measureText` call, no label-overflow probe).
+**Do not** put it in the runtime bundle: the app is offline by design, and the skill's own stack
+table imports it through a CDN, which contradicts that. The second use is one blog demo, not
+infrastructure.
+
+**mise (`jdx/mise`) — NOT NOW, with a named trigger that flips it.** Verified: mise-en-place, MIT,
+macOS/Linux/Windows, one CLI that declares tool versions, environment variables and commands in
+`mise.toml` and uses them in the shell, the editor and CI; the polyglot successor to
+asdf/nvm/pyenv, and `.tool-versions` already works; installed with `curl https://mise.run | sh`.
+The premise for adopting it did not survive measurement: the projects do **not** run differing Node
+versions, and nothing declares a version at all — 0 `.nvmrc`/`.node-version`/`.tool-versions` files
+under `~/projects`, `engines.node` in exactly **one** first-party manifest, and a single Node on
+the box (v22.23.1, reached through `~/.local/bin`, which a shell profile puts first). The one real
+divergence is a *package-manager* split, which mise cannot resolve. Adopting it now would add a
+second version source beside the Node Hermes bundles. **Adopt when either becomes true:** two
+projects need different Node majors, or the first Electron app lands (Electron ships its own
+Node/Chromium, so the host Node stops mattering for the shell and starts mattering for the build
+tooling). **The cheap thing to do meanwhile:** declare the version that already exists — one
+`engines.node` line where it is missing, and a `measured <date>` gate line in the HANDOFF naming
+the Node the gates ran under.
+
+44. **The version statement is a convention, not an enforcement row — the sync is test-side and
+    the engine itself never checks it.** W2 closed the measured hole PLAN-V1 §4.4 recorded (no
+    test read any of the five `GOBLIN_*_VERSION` constants in `bin/`): `tests/t-version-sync.sh`
+    now pins the count of those constants at 5, asserts every constant and
+    `package.json.version` equals `VERSION`, and asserts `goblin --version` — through both the
+    bash CLI and the node shim — prints `VERSION` byte-for-byte. What remains open: the sync
+    lives only in this repo's test suite, which a target repo never runs — and when the npm
+    tarball exists (W3/W5 packaging, planned to ship no `tests/`, deliberately), a published
+    package's `goblin.js` and its bash payload could
+    drift apart with nothing in the shipped artifact noticing. The engine has no self-check row
+    that reads its own `--version` against a manifest record, and adding one would make the
+    version a rule — which is a real option, not done here. Until then the guarantee is
+    development-time only: green in this checkout, unverifiable in the wild.
+
+45. **The migration is crash-safe by sequence, not by journal — a kill mid-upgrade leaves a shape
+    only three of which are detected.** `goblin upgrade` runs eight steps across two commits, and
+    the spec's refusal conditions name the shapes a crashed run can leave: record says global but
+    the 18 files are still here (R8 catches it), the record lost its engine: block entirely (R9
+    catches it), commits half-landed (the preflight's own state resolution). What is NOT caught:
+    a crash between the engine landing (step 3) and commit A (step 4) leaves `~/.goblin/engine`
+    written but the repo untouched — harmless, invisible, and never re-verified (the next upgrade
+    run compares hashes and refuses R6 if the payload has since changed, so the stale engine can
+    sit there being wrong until someone looks). And the shadowing footer (W3 §3) reads only the
+    vendored-payload shape; a repo whose engine_dir points at an engine that no longer exists
+    FAILs the resolution chain loudly (exit 2), but a repo pointing at an engine whose bytes have
+    silently changed since migration day runs green on the drifted table — the footer's hashes
+    name what ran, nothing compares them to migration day's. **Ticketed, not gated:** a
+    migration-day hash pin in the record compared per-run is the real fix and is a row-shaped
+    change (a new IN clause), not a W3 patch. Measured: R6 refuses the wrong-engine reuse, R8/R9
+    catch the two crash shapes they name, and the between-steps engine write is unwatched.
+
+46. **The emitted platform config is unsigned and hand-editable — DRIFT is detected per run,
+    never prevented.** `goblin emit` writes skills byte-copies and one delimited block in the
+    platform's context file, recorded in `~/.goblin-stack/emissions.tsv` with pre-image hashes
+    (so `--uninstall` restores byte-exactly), but nothing signs what it writes: an edit to an
+    emitted `SKILL.md` or to the bytes inside the `goblin-stack:begin/end` block makes the
+    platform's `verify.sh` oracle report DRIFT on the next `goblin doctor` run — and that is
+    all it does. There is no lock, no signature, and no write protection on any emitted file;
+    a platform (or the user) can change them between two doctor runs and nothing notices
+    until someone runs one. The same holds for `--unshadow`: it removes only project copies
+    whose hash equals the source payload and refuses-and-names any that differ, so a real
+    local edit survives, but nothing reconciles it either. Measured: a one-byte tamper in an
+    emitted `SKILL.md` and a stale marker VERSION in the context block each report DRIFT (exit
+    1) on the next run, and uninstall refuses to delete a recorded file whose bytes no longer
+    match its post-image (R6).
+
+47. **The four W4b adapter conventions are documented shapes, not run-probed installs — and
+    two of the seven platforms cannot block commands outright.** cursor and codex are not
+    installed on the build machine, so their rows pin the official docs (read 2026-09-29),
+    not a live CLI; a platform changing its layout invalidates the adapter silently until a
+    doctor DRIFT names it. codex and gemini report cap_command_blocking `partial` — codex
+    disables skills via `~/.codex/config.toml` `[[skills.config]]`, gemini only narrows via
+    approval modes — so an emitted skill is *available* there even when the operator would
+    forbid it; the doctor prints the codex hooks caveat (sessionStart only). And gemini's
+    id cell was measured to be exactly its platform name (the `gem_ini` typo shipped in one
+    W4b build and the doctor's DRIFT caught it — the schema check works).
+
+48. **`SP-01`'s rule text once overclaimed; it now states the check's actual scope.** The text
+    used to read "The current round has a SPEC" while the check is
+    `ls ./*-SPEC.md >/dev/null 2>&1`, which passes on ANY spec-shaped file at the repo root. The
+    measured consequence: a fresh class-A install ships the scaffold's `ROUND-000-SPEC.md`, and
+    `SP-01` PASSes on that scaffold alone — no round is open, and the row is green anyway. A stale
+    spec from a finished round keeps satisfying the row exactly as well as a live one, because the
+    check has no notion of "current". **Fixed in W6 (text, not check):** the row now reads
+    "A *-SPEC.md file exists at the repo root (any round, not the current one - round-scoping
+    arrives with the W6 staged chain)" — manifest and `docs/ENFORCEMENT.md` re-rendered in the
+    same commit, the check cell untouched. What it costs: the SPEC-exists signal in a run summary is
+    weaker than a round-scoped claim would suggest — read it as "a `*-SPEC.md` file is present", not "this round's spec is
+    here". Measured: `goblin-verify` on a fresh probe install reports `PASS SP-01 (ls
+    ./*-SPEC.md >/dev/null 2>&1)` with `ROUND-000-SPEC.md` the only file the glob sees, and the
+    same PASS after renaming it to a non-round name. **Ticketed, not gated:** "current round" is a stage-order notion, and the W6 staged workflow chain (SC-01:
+    a SPEC committed before the changes it governs) is what gives the word meaning; the staged
+    chain's stage-order rows are what will make round-scoping REAL — the reword names the
+    boundary until then.
+
+49. **The engine footer prints to captured stdout, and a script parsing `goblin-verify` output
+    must expect it.** The two-line footer (`engine: mode=… cli_sha256=… enforcement_tsv_sha256=…`)
+    is unconditional `printf` output — there is no isatty guard, by design, so a captured run still
+    names the engine that judged it (that is the point of #43's statement). The cost is parser
+    friction: a script that treats every stdout line as a verdict line trips on the banner and
+    footer lines, which are not `PASS`/`FAIL` rows. The contract is the EXIT CODE, never the line
+    set: 0 green, 1 red, 2 refuse — parse the exit code, or filter to `^[A-Z]{2}-[0-9]{2}` before
+    reading lines. Measured: `./.goblin/bin/goblin-verify > out.txt` on a probe install puts
+    `engine: mode=vendored cli_sha256=… enforcement_tsv_sha256=…` in the captured file
+    (`tests/t-engine-dir.sh` itself asserts the footer IN captured output, so removing it or
+    tty-gating it would break the engine's own suite). Recorded so the pitfall is findable from
+    this file; no change to the engine is implied or wanted.
+
+50. **`CM-01` reads only the most recent commit — historical commits with a wrong identity pass
+    unseen. Current stated scope: the row gates the identity of HEAD at the moment of the run,
+    and nothing older; that is the row's whole claim, by design.** The check is
+    `test "$(git log -1 --format='%ae')"` against the configured
+    `owner_email`, so it gates the identity of HEAD at the moment of the run and nothing older.
+    Measured: a probe history `owner → wrong@old.co → owner` reports `--only CM-01` clean (exit 0)
+    with the wrong-identity commit sitting one below HEAD. That is consistent with the
+    gate-at-the-moment design — every row judges the tree and history as they stand when verify
+    runs, and retrofitting an identity sweep over `git log --all` is a policy change, not a bug
+    fix — but it should be named: the row's green means "the latest commit carries the owner
+    identity", never "no commit in this repo's history carries an ambient one". A wrong-identity
+    commit that has since been followed by correct ones is invisible to every run. Recorded as a
+    boundary; no row change implied. `docs/ENFORCEMENT.md`'s CM-01 row carries the same one-line
+    scope statement.
+
+51. **`gob init` writes its three added values into `goblin.yaml` itself, not through a
+    template.** Install renders `.goblin/goblin.yaml` from `templates/goblin.yaml.tmpl` and
+    owns it (`put_once`: never rewritten after the first install) — and install correctly
+    carries no `--branch/--email/--gate` flags, because those keys are the project's to edit.
+    The wizard therefore `sed`-patches `branch:` and `owner_email:` and rewrites the first
+    gate's `cmd:` line right after install renders the file, fail-closed: the gate must read
+    back through the engine's own `g_yaml_gates` or the run stops. The cost is a second
+    writer for exactly those three lines: a hand-customised comment placement survives, but a
+    future template change to those lines' shapes (renamed keys, a multi-gate default) must
+    be mirrored in `bin/goblin-init`. The wizard never rewrites anything outside the declared
+    keys, and a re-run after a hand edit of another line leaves that line alone. Recorded as
+    a boundary; the alternative — teaching install three wizard-only flags — would put wizard
+    vocabulary into the installer's contract for no gain.
+
+52. **`scope:source` and `scope:target` name WHOSE burden a row carries — the framework's or
+    the adopting repo's.** `scope:source` is goblin-stack's own proof burden: those rows
+    (`PR-01`..`PR-05`) are the framework testing ITSELF while it is being developed — they run
+    in this repo, under `tests/run-tests.sh`, and a dev of goblin-stack is the one who owes the
+    run. `scope:target` is the adopting repo's proof burden: those rows run in an installed
+    repo via `goblin-verify`, and the repo's owner owes the run. Same matrix, two creditors:
+    a source row can never fail a user's repo, and a target row can never substitute for the
+    framework's own suite. Recorded as a definition; `docs/ENFORCEMENT.md`'s scope paragraph
+    carries the same sentence for the reader who arrives there first.