fix(devx): classify the tree-measurement claims under scripts/ instead of sweeping them

[os-try-charles]

Fixes #17797

The deliverable is a classifier, not a sweep. Three kinds of sentence live under one
phrase and take different repairs, and only one of them is a defect.

⭐ The card's own headline number had decayed — which is the card's best self-evidence

The card states that scripts/pm/dispatch-gates.mjs "alone carries 40 sentences
containing measured on this tree". Re-measured at 0046a41b4, 2026-09-20T00:30Z, with
the card's own case-sensitive probe: 14, of which 3 carry a revision by the card's
own rev probe. A card about bare magnitudes decaying in silence had its own bare magnitude
decay by 65% before anyone read it. ⛔ The 40 is not carried anywhere in this PR.

1. The declared population — which files, which phrase forms

All readings below are mine, taken at 0046a41b4 (this branch's merge base),
2026-09-20T00:30Z. ⚠️ Every one of them decays; the tool added here re-derives them.

probe	occurrences	files
the card's probe — measured on this tree, case-sensitive, over scripts/	46	25
all three spellings that actually occur, case-insensitive — measured on this tree, re-measured on this tree, measured against this tree	121	47
the same, excluding the two files scoped out below	74	45

⚠️ That is an upper bound on a population, ⛔ not a count of defects. The card says so
in its own words and this PR does not contradict it.

Three findings about the population itself, each of which says a phrase probe is not a
population:

• The card's probe is case-folded shut. Lower-case-only reaches 46 of the 121 sentences
  that carry the phrase in some spelling. The gap is not rounding.
• A line-based probe cannot see a sentence that wraps. The phrase itself straddles a
  line break in five files, so no grep of it lists them at all:
  scripts/check-examples-live-imports.mjs, scripts/check-parse-guard.mjs,
  scripts/check-self-test-wired.mjs, scripts/eslint-fatal-guard.mjs,
  scripts/i18n-bundle-surface.mjs. The tool reads sentences, so it holds them.
• ⭐ And the card's own named instance carries no phrase at all. It is a bare corpus
  size in the present tense. A population defined by the phrase would have reported the
  tree clean at the one site everybody already agreed was broken — so the population is
  declared in two forms, FORM A (the phrase) and FORM B (declared sites the phrase cannot
  reach, currently the hand-written-corpus size in scripts/docs-audit/README.md).

2. The classifier

scripts/pm/measurement-claim-triage.mjs:

node scripts/pm/measurement-claim-triage.mjs              the population + every verdict
node scripts/pm/measurement-claim-triage.mjs --review     the kind-3 residue alone
node scripts/pm/measurement-claim-triage.mjs --self-test  the controls, both directions

It records no number about its own population — the scope line it prints is the number,
re-derived on the tree it runs against. That is the durable half of the fix: the reason the
card's 40 went stale is that it was written down instead of re-derivable.

Four tests, applied in that order. Each can only move a hit out of the residue, and each
prints the cue that moved it, so every verdict is auditable against the sentence.

• T1 FRAME — is the magnitude scoped to a moment that has passed, or to a counterfactual
  condition? (before this gate was written, at the commit that took it, an earlier revision, with the reservation removed, was auditing) ⇒ KIND 1.
• T2 ANCHOR — does the sentence carry a revision or a date? ⇒ KIND 1.
  ⛔ A bare card number is deliberately not an anchor: #NNNN occurs in every kind of
  sentence here, and admitting it would clear the residue by matching everything.
• T3 LOUDNESS — would the drift announce itself? Three ways: a named failure direction
  (a zero whose falsifier is spelled out is the commonest), a pointer at the instrument that
  reprints the value, or a live assertion that reds on drift. ⇒ KIND 2.
• T0 MAGNITUDE — last, and weakest: nothing decays if nothing is counted. A ZERO counts,
  because kind 2 is defined on one.
• Everything else is REVIEW, a candidate bucket. ⛔ Reading its size as a finding count
  is the error the card exists to prevent.

T4 — level or relation — is a judgement and no regex decides it. Does a reader act on
the magnitude's level, or on a relation (a ratio, an ordering, an existence claim, a set
equality) that the level's drift survives? A level is kind 3; a relation is an
ILLUSTRATION, and rewriting it buys nothing. T4 is recorded per site in the tool's TRIAGE
table with its reason, and a REVIEW hit with no row prints UNTRIAGED and reds the self-test.

⛔ Every TRIAGE row is keyed by an excerpt, never by a line number — the card cited its
own known instance at :493 and it was at :555 by the time anyone read it.

The window, and a bug worth recording

The first build read a fixed ±6-line block and cleared the one site everybody had agreed
was broken: six lines above it an unrelated sentence cites a revision, and a block window
handed that anchor to a claim that has none. T2 and T3 now read the sentence; T1 also
sees the sentence before (a narrative frame leads), T3 also the sentence after (a named
failure direction trails — the card's own kind-2 exemplar spells the zero in one sentence
and names its falsifier in the next); ⛔ T2 sees neither. That control is pinned in the
self-test.

3. ⭐ Controls, both directions

node scripts/pm/measurement-claim-triage.mjs --self-test → green, six controls:

control	site	expects
kind 1 — a citation carrying its revision, left alone	scripts/check-tier-file-adoption.mjs ("Measured on this tree at d03c3c96d6 …")	KIND-1
kind 2 — a zero whose failure direction is named, left alone	scripts/docs-audit/affected-docs.mjs ("zero commands declare static topic …" + "The failure direction if that ever changes …")	KIND-2
kind 2 — a refusal a live assertion holds, left alone	scripts/check-skill-compatibility-version.mjs ("the refusal is pinned in the self-test …")	KIND-2
kind 3 — the card's named instance, CAUGHT	scripts/docs-audit/README.md (pre-repair text)	REVIEW
kind 3 repaired — the same site now points at the gate	scripts/docs-audit/README.md (post-repair text)	KIND-2
⛔ the one-sentence anchor rule — a neighbour's revision does not clear a bare magnitude	synthetic	REVIEW

Each control carries a liveness key that must still be present in the file it was quoted
from (and the pre-repair one carries an absent key instead — if that sentence comes back,
so has the defect). ⛔ A control quoting a sentence the tree no longer holds passes in
silence, which is the failure this whole PR is about. The self-test also reds on a dead
cue: every cue must still match something in the population or the controls — nine that
matched nothing were deleted rather than left as decoration.

4. The kind-3 hits and their repairs

Five, judged out of a pre-repair review residue of 27. After the repairs the tool reports
87 claim(s) over 48 file(s) — KIND-1 35 · KIND-2 11 · NO-FIGURE 19 · REVIEW 22, and all 22
survivors are recorded ILLUSTRATION. ⛔ No fresh bare number is written anywhere in this diff.

Repaired by pointing at the live instrument (#16200's preferred shape):

• scripts/docs-audit/README.md — "the anchor derivation reads the same 178-page corpus
  the old one did". Present tense, and wrong today: pnpm check:docs-audit-scope printed
  195 hand-written doc(s) at 2026-09-20T00:11:30Z, my own reading. Repaired to name the
  corpus without sizing it and to send the reader to that gate, and it says out loud that
  the figure is deliberately absent.

Repaired by pinning the reading to the commit that recorded it — the historical figure
does work a live reading cannot (it justifies a decision taken at that moment) and no gate
reprints it. Each carries a deliberately-absent note, so the next author does not helpfully
restore a present-tense figure:

• scripts/check-pnpm-filter-targets.mjs — the scripts/** declaration-honesty ratio, now
  Measured at 52a41b72e (2026-08-23). The counts move with every file added under
  scripts/; the SHARE is what the paragraph argues.
• scripts/check-published-files.mjs — the same shape over the whole publishable workspace,
  now Measured at 52a41b72e (2026-08-23).
• scripts/check-skill-compatibility-version.mjs — the four precision ratios deciding which
  roots are declared, now measured at f29e89717 (2026-08-22), with a pointer to
  scripts/pm/bare-root-worklist.mjs, which carries this gate's package-root rows with
  their own dates.
• scripts/check-slot-lookup-ratchet.mjs — 46s, the worst-ageing kind of figure: a
  wall-clock reading is a reading of one box as much as of one tree, so it decays without
  the tree moving at all. Now Measured at 46s when this was recorded (99ca6623f, 2026-08-19), with the order (seconds against a CI round) left as the load-bearing claim.

Each repair's excerpt is recorded REPAIRED in TRIAGE, and the self-test reds if it
returns to the population.

⛔ What the classifier deliberately did NOT touch

The 22 remaining REVIEW hits are recorded ILLUSTRATION with a reason each. Two worth naming,
because they are the shape a sweep would have destroyed:

• scripts/docs-audit/README.md lines carrying 178 in the historical narrative — the card
  predicted these were kind 1 and judging each one agrees. Two more, at the --all backstop
  and the cost note, already point at check-audit-scope.mjs in the same sentence and
  classify KIND-2 mechanically: already repaired in [finding] the docs-audit union-read prose hard-codes the anchor lag as "532 keys" in two places — measured 604 four days later #16200's shape, so ⛔ this PR leaves
  them exactly as they are.
• scripts/check-dispatcher-error-vocabulary.mjs's per-glob member counts: kind 2, not
  kind 3 — the docblock names what would falsify them and PUBLISHED_SOURCE_FACE_FLOOR
  plus the per-glob presence pin red in that direction.

⛔ Two files are outside the file surface, for two independent reasons

scripts/pm/dispatch-gates.mjs and scripts/pm/check-widening-tells.mjs are not edited
here, and both reasons are carried as data in the tool's EXCLUDED table so every run
prints them rather than leaving them in prose:

1. Live conflict. dispatch-gates.mjs is being edited by open PRs ci: publish the platform checklist status pages to the repository wiki on a schedule #19162 and
   revert(spec): take back the declaration-text snapshot, restore the 27 signature hashes #19024; check-widening-tells.mjs by fix(pm): check-widening-tells reads the member’s PRIOR schema — a bound inside a previously-z.unknown() bag is not a new key #19153 and revert(spec): take back the declaration-text snapshot, restore the 27 signature hashes #19024. Editing underneath
   them is a conflict, not duplicated work.
2. The card's own verification constraint: scripts/pm/dispatch-gates.mjs is enormous
   and its --self-test exceeds the agent container's foreground cap (recorded on [finding] dispatch-gates.mjs --self-test asserts an empty child stderr too, so it reds under --use-env-proxy exactly as #15234 did #17765),
   so the card required whoever took this to say how they verified a change to that file
   or scope it out and say so. This is the saying-so.

⇒ The residue those two hold is un-swept, and the tool's scope line says so on every run
instead of reading as complete.

Also untouched, and reported rather than edited: .claude/**, skills/**, docs/adr/**,
AGENTS.md, CLAUDE.md (governed surfaces) and #16200's three carriers, settled by #17795.

Verification

Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack
at 60f9b648a — 6 paths, 722 changed lines. 35 derived families, 35 run, all exit 0;
reconciled with --ran: 35 derived famil(ies) accounted for — 35 run, 0 UNRUN.

• node scripts/pm/measurement-claim-triage.mjs --self-test → ✓ controls hold in BOTH directions (6 controls, 27 recorded judgements).
• pnpm check:pm-dispatch-gates → ✓ dispatch-gates self-test: 1866 cases pass,
  593.2s on this box, VERDICT command-exit 0, run detached under
  scripts/pm/os-verify-lock.sh (held 594s, waited 0s), 2026-09-20T00:23:16Z–00:33:10Z.
  ⚠️ The reference line at content/docs/qa/platform-readings.md:432 records 430–450s; this
  container reads well above that band, and this run is one more point in it.
• ESLint, narrowed and the narrowing measured: npx eslint --no-inline-config --format json
  over the five changed .mjs files at 60f9b648a → 5 files linted, 0 errors, 0
  warnings (counts read from the JSON reporter). The narrowing excludes nothing:
  eslint.config.mjs enables type-aware linting for no configuration in this repo (no
  parserOptions.project, no typed rules), so this diff cannot move the verdict on a file
  it does not touch. The repo-wide eslint . run is CI's.
• ⊘ NOT MEASURED — pnpm check:published-readme-exports and pnpm check:dts-closure
  both exited 3, PREREQUISITE NOT MET (46 packages' dist/ not built). Neither is in
  the derived family; both are artifact-roster families whose roster sits under scripts/,
  run here only because silence there is not evidence. This diff changes no package source,
  so neither could be moved by it. ⛔ Recorded as not measured, not as a pass.
• No changeset: measured, not assumed. The root manifest is private: true and no package's
  files[] names the repo-root scripts/ tree, so nothing published moves. Labelled
  skip-changeset.

Acceptance notes (out of scope, noted and not filed)

• scripts/pm/dispatch-gates.mjs holds 42 of the 121 phrase occurrences (all spellings) and
  is un-swept here for the two reasons above. Carrier: the PR that next touches that file
  — ci: publish the platform checklist status pages to the repository wiki on a schedule #19162 or revert(spec): take back the declaration-text snapshot, restore the 27 signature hashes #19024. Worth a card of its own once they land; the tool classifies it the
  moment its EXCLUDED row is removed.
• scripts/pm/check-widening-tells.mjs holds 5, same disposition. Carrier: fix(pm): check-widening-tells reads the member’s PRIOR schema — a bound inside a previously-z.unknown() bag is not a new key #19153 /
  revert(spec): take back the declaration-text snapshot, restore the 27 signature hashes #19024.
• The tool is not CI-wired: a package.json script and a workflow step both sit outside
  this card's file surface, so --self-test is run by hand and its controls watch nothing on
  their own. The docblock says so out loud. ⛔ Do not read a green run here as CI coverage.
  Carrier: none today — it needs a seat that owns the root manifest and the lint
  workflow. Raised as an open question in the report rather than filed, because wiring it is
  a decision about CI cost, not a defect.

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

Nothing in this diff resolved to a documentable surface (no symbol, route or SDK anchor derived from 0 changed package(s)), so this run has no opinion about the docs.

What this run could not see

• a page that states a rule by its inputs shares no identifier with the emitter that implements the rule, so an emitter-only diff cannot list it — not on this run and not on any run. Measured on fix(driver-sql): emit varchar(maxLength) for a text field a declared index keys on #11430: content/docs/protocol/objectql/types.mdx documents the text-family column mapping by the ObjectQL type names it maps FROM (text / textarea / html) while the diff changed createColumn; it went unlisted, and it was the page that diff falsified, in four places. No shared token exists to detect this on, so a rule your change carries has to be re-read by hand in the pages that restate it.
• a key NAME is not a key, so the hand re-read the line above prescribes can land on the wrong schema. The same spelling is authorable on one governed type and a [REMOVED] tombstone on another for each of active, aria, joins, objects, template, tools and version (censused on [finding] tools is a key on BOTH AgentSchema (tombstoned, dead) and SkillSchema (live, cloud-attested), so a name-based search attributes skill examples to the agent key — it produced a false stop-the-line alarm on PR #19059 #19093 over the liveness ledger's governed types, top-level keys); nothing in a search result distinguishes the two, so a grep hit on a LIVE example reads as evidence about the DEAD key. Measured on fix(spec): the agent.tools liveness row says dead — it claimed live on a key the schema tombstoned #19059: content/docs/ai/agents.mdx was reported as contradicting the agent.tools tombstone over its tools: example at :161, which is inside the defineSkill({ block opened at :155 — the page was already correct. Settle ownership by PARSING the value against both schemas, never by the name: that literal PASSES SkillSchema, and as an AgentSchema it FAILS at tools with the tombstone prescription. ⛔ These names are not the whole class — a key retired through a .strict() guidance map leaves no tombstone in the walked shape and none of them here (tool.category, live as AIToolDefinition.category).

Coarse fallback — 0 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): node scripts/docs-audit/affected-docs.mjs --json 0046a41b43da8e7d2db9d7f926cc0518a357c379 → packageMentionDocs.