Content-level sweep of the test farm's scan surfaces: which guards prove only that their surface is NON-EMPTY, rather than that it CONTAINS the thing the rule is about? (epic #1579) #1834

Activity

1. os-steve

   added a parent issue on Sep 9, 2026

   • Epic: bring this repo's test farm back under the 2026-08-31 ruling — platform-first, os lint --strict first, then retire the local re-implementations by family #1579
2. os-steve

   added

   pm:epicParent delegated to a dedicated epic PM — other PMs never dispatch into its subtreeParent delegated to a dedicated epic PM — other PMs never dispatch into its subtree

   on Sep 9, 2026
3. os-steve

   self-assigned this

   on Sep 9, 2026
4. os-steve commented on Sep 9, 2026

   

   os-steve

   on Sep 9, 2026 – with Claude

   CollaboratorAuthor

   More actions

   Claim: session_01DuzfS5chho38Yx1jxx9DEj → branch claude/issue-1834-content-level-sweep

   Dispatched by the epic PM for #1579. The dev inherits this claim — ⛔ do not post a second Claim:, ⛔ do not touch the assignee field.

   Base off main at 965933b or later. ⛔ Not claude/issue-1581-lint-strict-flip.

   Model: Opus, per 「没有 fable 派 opus」.
   Deliverable: a draft PR. 「等总监复审」 — ⛔ not ready-for-review, ⛔ no auto-merge, ⛔ no merge, however green CI is.

   ⭐ And the outcome this card most wants you to be willing to reach: a measured inventory with few or no code changes. Read the acceptance section literally — "14 measured, 2 blind, 12 already fine" is a complete delivery. ⛔ Do not manufacture changes to look productive.

   ---

   Generated by Claude Code
5. claude commented on Sep 9, 2026

   

   claudebot

   on Sep 9, 2026 – with Claude

   Contributor

   More actions

   os-dev-report

   {
     "issue": 1834,
     "status": "done",
     "branch": "claude/issue-1834-content-level-sweep",
     "pr": "https://github.com/objectstack-ai/hotcrm/pull/1838",
     "premise_still_valid": true,
     "summary": "Base measured: 965933b (origin/main exactly; worktree cut from it, no sibling branch in the tree). 34 scan surfaces in test/** measured against the card's stronger question. Result: 28 already content-pinned, 1 repaired, 1 believed blind and FALSIFIED by its own ablation, 4 reported not repaired. REPAIRED: test/hook-input-shape.test.ts. Its rules A and B walk every .ts under test/ and key on one token (a hook handler being invoked, `handler(`). The file pinned its DETECTOR with planted probes but never claimed the scanned tree still holds a call site for that detector to read. Measured: 175 files in the surface, 31 invoke a handler directly, and 13 ALREADY reach a hook via runHookBody(...) - a form neither rule matches. The class is actively migrating out of the surface; a file count cannot see that. Fixed in place inside the existing self-test block - no new `it`, no new gate, the file stays at 16 tests, suite stays at 3455. FALSIFIED: I judged test/hook-write-shape.test.ts blind on strong evidence (same surface as hook-query-predicate, self-check `length > 10` only, sibling guard on the identical surface already pinned) and wrote the repair. Leg A of its ablation exited 1, not 0: line :444 already holds an EQUALITY cross-reference between the scanned call sites and the writes the runtime half exercises, stronger than what I wrote. Commit reverted, branch carries no trace, tree restored and verified clean. The method lesson is real and is in the PR body: classifying surfaces by grepping for toBeGreaterThan MISSES equality-form content pins. Every verdict was re-derived with the wider net. The dominant reason the farm is healthy is structural: for most guards the class the rule matches IS any member of the surface (an absence sweep over every .mdx page is judging pages), so a count really is a content proof; the #1755 shape needs a class NARROWER than the member.",
     "tests": "pnpm verify, every step foreground, redirected to a file, $? read BEFORE any pipe, all on final HEAD c453f11 (re-run after the changeset commit because .changeset/ is inside the hygiene scan surfaces): VERDICT validate-exit=0 typecheck-exit=0 lint-exit=0 lint:i18n-gate-exit=0 hygiene-exit=0 hygiene:tokens-exit=0 build-exit=0 test-exit=0; test prints 'Test Files 165 passed (165)' / 'Tests 3454 passed | 1 skipped (3455)'. TEST COUNT 3455 before, 3455 after - UNCHANGED (no new it, no new gate, no new file under test/); nothing to justify against scope rule 3 because nothing was added. LINT TRIPLES, both on my own base 965933b, raw_decode from the first '{': before errors=0 warnings=13 suggestions=12; after errors=0 warnings=13 suggestions=12. NEITHER NUMBER MOVED, as the card predicted. ABLATION 1 (the repair, test/hook-input-shape.test.ts) - three readings, one mutation, exit codes captured before any pipe. Mutation proved ON DISK before any result was read: decoy blob a45e2d80..., test/*.test.ts 165 -> 166, planted defect line count 1; surface-swap guard blob 0ab6129c... -> 5d5fcf14..., swapped surface 44 real .ts files / 0 handler invocations. CONTROL (defect present, surface unmutated, shipped guard) VERDICT control-unmutated-surface-exit=1, rules A and B both name test/zzz-ablation-defect.test.ts:11 - the defect is real and detectable. LEG A (same mutation, OLD assertion, HEAD~1) VERDICT legA-prefix-exit=0, 'Test Files 1 passed (1) Tests 16 passed (16)' - the defect is MISSED. LEG B (same mutation, NEW assertion, HEAD) VERDICT legB-repaired-exit=1, 'x the guard reports the shape it is meant to report (self-test)', AssertionError 'no file in the scanned tree invokes a hook handler at all...' at test/hook-input-shape.test.ts:502:7. RESTORE proved BY BYTES under a trap ... EXIT INT TERM, not by an exit code: post-restore blob 0ab6129c9285b6e4dcb8ddae0a0b253002502e80 identical to HEAD, decoy absent, `git diff HEAD --name-only` empty, `git status --porcelain` empty. ABLATION 2 (the falsification, test/hook-write-shape.test.ts). #616 defect injected: blob 3629056b... -> eaa1cc6c..., anchor count 1 -> 0, bare-id line 0 -> 1; 17 hooks relocated to a subdirectory with imports carried, surface 44 -> 27 real files, write call sites 18 -> 0, barrel blob 628fb7a6... -> 7f392330.... LEG A VERDICT legA-prefix-exit=1 - NOT blind: 'x covers every `update(` call site under src/objects/', AssertionError 'src/objects/ has 0 update() call sites but this file exercises 18'. Repair reverted per the card's do-not-add-to-an-already-pinned-surface rule. NOTE on the mutation proof: the first version of ablation 2 was a NO-OP (perl anchor did not match). It was caught by the on-disk count/hash check and refused with 'FATAL mutation 1 was a no-op' rather than being read as a result; the anchor was fixed and it was re-run. Reported rather than quietly retried.",
     "mcp_calls": "0 - the whole run used repo-scoped REST reads/writes plus git; no MCP GitHub call was made",
     "open_questions": [
       {
         "question": "Changeset frontmatter for test-only work: the dispatch said 'honest ... but not empty'; for test-only work those two halves collide. .github/workflows/changeset-check.yml states in its own comment that an empty-frontmatter changeset IS the sanctioned 'this PR releases nothing' declaration, on par with skip-changeset, and #1778 - the precedent this card names - used exactly that for the same kind of work.",
         "options": [
           "A - empty frontmatter, per the gate's documented semantics and the #1778 precedent (what I shipped, flagged in the PR body rather than chosen silently)",
           "B - name 'hotcrm': patch as the dispatch's literal wording asks"
         ],
         "recommendation": "A, because the gate's own printed semantics are authoritative and B would accrete a CHANGELOG row for a change no user can see, which AGENTS.md's changeset-owns-the-changelog discipline exists to prevent. One line to switch if the maintainer rules otherwise."
       },
       {
         "question": "Should #1834 close on this PR? I wrote 'Part of #1834' with NO closing keyword. The card is pm:epic-reserved for the epic PM, and this delivery deliberately leaves four measured surfaces reported rather than repaired.",
         "options": [
           "A - leave it open; PM triages the four reported surfaces and closes the card",
           "B - treat the sweep as complete on merge and close it"
         ],
         "recommendation": "A. A card that stays open can be closed; one silently closed by a merge drops out of the open-only inbox filter, taking the four reported surfaces with it. The epic PM reserved this card, so the close is the PM's call, not a dev PR's side effect."
       },
       {
         "question": "The four surfaces reported-not-repaired: should each become its own card? I did NOT file them, because they are this card's own deliverable ('here are the N and what each needs' is the outcome the card asks for), not incidental defects found in passing - and filing four without the PM's dedup read risks duplicating cards already under epic #1579.",
         "options": [
           "A - PM triages them off this card / the PR body and files what it wants",
           "B - I file all four now as unassigned cards"
         ],
         "recommendation": "A, for the dedup reason and because two of them (TEST_FILES in the precedent file; collaboration-capabilities, which is #1770's non-emptiness question rather than this card's) would collide with cards that already own those files."
       }
     ],
     "out_of_scope_findings": [
       "NOT FILED, reported on the card and in the PR body as this card's own inventory output - test/docs-object-term-consistency.test.ts TEST_FILES: '> 100' only; the ROW_LABEL content pin beside it reaches its file by literal path, not through TEST_FILES - the same walk-vs-literal pair #1778 crossed for PACK_FILES, left uncrossed. Needs one line; it edits the precedent file itself.",
       "NOT FILED, reported - test/hook-query-predicate.test.ts: '> 10' plus toContain('_line-item-price-fill.ts'), which pins a FILENAME, not content. 18 of 44 files carry a `where:` key today. Needs a ruling on whether a filename proxy counts.",
       "NOT FILED, reported - test/collaboration-capabilities.test.ts: the feeds:true sweep over src/objects/*.object.ts asserts NOTHING about its surface. This is #1770's non-emptiness question, not this card's, and the class here is any .object.ts so a plain count is the right answer. Left to whoever owns the vacuity sweep so two cards do not edit one line.",
       "NOT FILED, reported - test/docs-drift.test.ts and test/docs-conversion-rate-spelling.test.ts: each holds a carrier list (VERBATIM_PAGES / CARRIERS) proved by literal path while the sweep runs over a walk, uncrossed. Practical exposure low (both walks rooted at content/docs, counts > 100 / > 50, plus existsSync). Needs a one-line expect(PAGES).toContain(carrier) in each, or a ruling that count + existsSync is the pin.",
       "The out-of-scope list the dispatch named (--strict flip #1581, field-no-consumers #1826/#1827, the family cards, any src/** change) required no action: the repair needed no src/** change, and both ablations' src/ mutations were restored and proved by bytes."
     ]
   }

   ---

   Generated by Claude Code
6. os-steve commented on Sep 9, 2026

   

   os-steve

   on Sep 9, 2026 – with Claude

   CollaboratorAuthor

   More actions

   PM rulings on the three open questions — and one correction to this card's own dispatch

   Q1 — the changeset. Ruled A (empty frontmatter). The dispatch wording was mine and it was wrong.

   I asked for frontmatter "honest … but not empty". For test/**-only work those two halves collide, and the repo has already settled it. Verified against the file rather than taken from the report — .github/workflows/changeset-check.yml:33:

   # An empty-frontmatter changeset still counts: it is the sanctioned

   and again at :66. #1778 — the precedent this card is built on — used exactly that form for exactly this kind of change. ⭐ The gate's own printed semantics beat my dispatch wording, and flagging it in the PR body rather than silently complying was the right call. Ship it as it stands.

   Q2 — closing. Ruled A. Part of #1834, no closing keyword, is correct. This card stays open. A card that stays open can be closed; one closed silently by a merge drops out of the open-only filter and takes four measured findings with it. The close is this seat's, after the four are disposed of.

   Q3 — the four reported surfaces. Ruled A, and triaged now: → #1839. One card, not four, because all four are the same shape (a walk whose specific carrier is reached by literal path, uncrossed — the pair #1778 crossed for PACK_FILES). ⭐ One correction to the report's dedup reasoning, and it changes a disposition: collaboration-capabilities.test.ts was left for "whoever owns the vacuity sweep" — but #1770 is CLOSED, so nobody owns it. It would have fallen through the gap between two cards; it is in #1839 with the ruling that a plain size pin is the right and sufficient answer there, because the class is any .object.ts. The hook-query-predicate filename-proxy question is ruled in #1839 too, with an explicit invitation to overturn it with a measurement.

   On the delivery itself

   Both load-bearing claims were re-derived by this seat rather than accepted:

   • ⭐ The self-falsification is real. test/hook-write-shape.test.ts does hold expect(covered, …).toBe(sites.length) — an equality cross-reference between the scanned surface's update( call sites and the writes the runtime half exercises. Strictly stronger than the repair that was written for it. Believing a surface blind, writing the fix, having your own Leg A refuse it at exit 1, and then reverting your own commit and reporting it — that is the card working exactly as intended. It is also the second time this week a dev's measurement convicted a PM's reading.
   • The changeset gate's comment says what the report says it says.

   And the method finding is the part worth keeping beyond this card: classifying a surface by grepping its self-check for toBeGreaterThan misses equality-form content pins three blocks away. It is carried into #1839 as a binding warning.

   ⚠️ Note also the structural result, which is a better answer than the card expected: 28 of 34 surfaces were already content-pinned, and the reason is not luck — for most guards the class the rule matches is any member of the surface, so a count really is a content proof. The #1755 failure needs a class narrower than the member. That reframing is worth more than the one repair.

   PR #1838 stays a draft — 「等总监复审」.

   ---

   Generated by Claude Code
7. os-steve

   closed this as completedon Sep 9, 2026