Skip to content

Batch 2026-11-21q: ADRs 1172 and 1167 (a store written through a prop; a never-rerunning effect's returned cleanup) - #315

Merged
yumemi-thomas merged 8 commits into
mainfrom
claude/batch-2026-11-21q
Nov 21, 2026
Merged

yumemi-thomas merged 8 commits into
mainfrom
claude/batch-2026-11-21q

Conversation

@yumemi-thomas

@yumemi-thomas yumemi-thomas commented Nov 21, 2026 •

Copy link
Copy Markdown
Owner

Bundle q (Mac 24c), from main fdc28ba (#314). Two accepted items, each merged normally on its own base (neither carries foreign history), and my own commits: ADR 1172 (two commits: the fail-closed fix for a silent clean on main, then R4 N1) and ADR 1167 (two commits). The authors ran neither make verify, make corpus-delta nor the fuzz smoke (1172 ran make verify-delta and the ownership gate): this bundle runs them, the census and the rest leg of the smoke included. ADR 1166 was in the first build of this bundle and is out (below). ADRs 1164/1165, 1177 and 1178 are not in.

In the bundle (each accepted by the coordinator on evidence)

  • ADR 1172 (claude/prop-store-write-1172 286b374, two commits on 14469fa).
    • Part 1 (8986792), a fail-closed fix for a silent clean on main: a store handed by prop and held in a const alias (const held = props.s1; held.a.v = 9) had no row of any rule at the Get held.a, though Chrome warns STRICT_READ_UNTRACKED there (runtime-oracle/prop-alias-link-get: unproven-clean on main). value_use::Census::member counts a decided write only where it is the member's own Set; a deeper write leaves the Get open as SC9012 value-held-in-binding.
    • Part 2 (286b374), R4 group N1: props.s1.a.v = 7 with <C1 s1={s1} setS1={setS1} /> and s1 a local createStore was uncertifiable (store-write-unproven); ADR 0491's literal proof now accepts a hand-off to an exact local component (handed_component_prop, component_prop_reads, const_alias_reads, handed_container_write). Author's measure: R4's 1,441 N1 programs plus 396 legal STORE_WRITE programs, U -> V 428 (mx-src 208, mx-pairs 208, mx-combo 12), no legal row moved, no V lost.
  • ADR 1167 (claude/effect-cleanup-1167 71968b7, two commits on b06a4e7): a never-rerunning effect's returned cleanup runs only at disposal, so a read written in its body is placed as an onCleanup callback's (write_roles::returned_cleanup_disposal_only, reads only). Author's probe (1,551 recall programs in the effect-cleanup-* and render-cleanup-* contexts): lg-recall 100 uncertifiable -> clean, nothing else moved.
  • Facts schema stays 108 (neither touches it).

Held out: ADR 1166

It was in the first build (claude/named-callback-1166 e93b741) and its moves held up against Chrome (3,946 U -> V and 2 U -> clean on the smoke's rest cover, judged with --runtime-diff all: 3,943 + 3 agree and 2 clean, 0 false-positive, 0 unexecuted). But the full census smoke failed with two false-cleans that it alone causes: mx-resolve-in-reactive-scope-barm-carried-body-x-direct-none and mx-until-in-reactive-scope-barm-carried-body-x-direct-none (createEffect(f1, ...) plus [f1].forEach((g) => g()), f1 calling resolve or until): main has an uncertifiable function-run-unplaced row there, with 1166 the row is gone and nothing names the rule (Chrome: "Cannot call resolve inside a reactive scope"). Main + 1172 keeps the row, main + 1172 + 1166 drops it, ablating its reporting-run arm does not restore it. Ruling of the coordinator: drop it, back to its author with mac24/q-1166-false-clean.md. 1167 does not depend on it.

Moves, read against main fdc28ba

  • Coverage (make coverage: nothing moved after the merge; every snapshot change of an existing fixture is a named one). ADR 1172: store-handed-root Nested SC2003 uncertifiable -> violation; local-alias-member-reads +2 and store-write-literal-links +5 SC9012 value-held-in-binding rows (the fail-closed rises, named in rises.json); new fixtures prop-alias-link-get and store-write-through-prop (4 violations, 4 uncertifiable). ADR 1167: SC1001 uncertifiable rows go in async-read-under-jsx (1, CleanupInShow) and effect-cleanup-rerun-read (2, NeverReruns); new fixtures effect-cleanup-disposal-only and effect-cleanup-disposal-reachable.
  • Checker-only scan of the smoke's rest-leg cover (blocker-scan.mjs, no Chrome: the 45,619 programs of the cover drawn as CI draws it, minus the census grid; main's release checker against this build's; each program classified by its site and by the uncertifiable kinds at the site, so a row lost under other remaining rows shows): 45,595 identical, 24 changed: 3 mx-pairs STORE_WRITE_OUTSIDE_SETTER U -> V (ADR 1172 part 2), 1 lg-recall effect-cleanup-for U -> clean (ADR 1167), and 20 programs that gain one SC9012 value-held-in-binding row at the site (class kept: 16 U, 4 V; STORE_WRITE programs: ADR 1172 part 1's fail-closed fix). No row is lost anywhere.
  • Monotone, main's release checker against this build's (textbook and fixtures, 1,785 programs): 0 errors, 11 (site, code) moved up, M2 four programs with a downward move, all SC9012 clean -> U of ADR 1172 part 1: local-alias-member-reads (2), prop-alias-link-get (2), store-write-literal-links (3) and store-write-through-prop (2; a new fixture of part 2, its alias Gets). No violation moves down.

Held out as well

ADRs 1164/1165 (W8), 1177 and 1178 wait for the coordinator's go. Nothing unaccepted is in this branch.

My commits

  • Merges of ADR 1172 (602bce8) and ADR 1167 (8ca6230); counts.json (ac6bab5): main plus each branch's own delta against its own base, additive, and the branches' own drops rows (1172: 1, 1167: 2); count.mjs passes without --update (308 keys, 37,502 rows, 481 drops).
  • boundary gap (f6fb505): ADR 1167's execution_role.read_execution_role.returned-cleanup-disposal is a silent-capable arm with neither a witness pair nor a stated gap, so scripts/fuzz/boundary.test.mjs fails on the branch and on any join; one gap is stated, naming the ADR's own fixtures, Chrome cases and probe. rises.json needs nothing of mine here: 1172's two rises came with its branch (rises: 2 measured, 2 named).

Merge notes

  • ADR 1172: conflicts in docs/open-work.md (two notes added at one place: kept both) and scripts/guard/rises.json (both sides appended: ours plus the branch's two rows, by a JSON-aware resolver); decisions.md, .gitignore, rediscovery.json and counts.json by my merge helper. No Rust conflict. ADR 1167: no conflict (it conflicted only with 1166, in read_execution_role).

Checks (this Mac)

Everything below ran on the PR head f6fb505, except where make verify says otherwise.

  • cargo check --workspace --all-targets: clean. make coverage: 1,764 fixture projects, 40,375 findings, nothing moved.
  • make textbook-check: 21 of 21 certified, 0 uncertifiable (site, code) keys, 0 newly proved, 0 problems.
  • CI's scripts-and-cli steps, run here (sh -n scripts/*.sh, vitest on scripts/*.test.mjs, the fuzz test lane): pass, 1,940 fuzz tests, 0 failed, 0 skipped (the boundary test included).
  • The full census smoke leg, CI's exact command: rc 0, cover 90,191 programs (sha256 87c941dff59d047f), 44,572 judged, 0 worse, 209 better or changed, proved verdicts 7,041 in the ledger, 4 lost (4 named), 715 newly proved, 11 known-open programs still failing (8 unplaced, 3 false-positive; open-defects.json is not touched), nothing needs a look.
  • The full rest leg of the smoke, CI's exact command (both shards as one job): rc 0, 45,619 programs judged (mx-recall 34,298, lg-recall 6,552, mx-pairs 2,771, lg-pairs 1,121, mx-src 365, lg-src 182, mx-base 172, mx-combo 100, lg-base 29, lg-combo 29), 0 worse, nothing needs a look; proved verdicts lost: mx-recall 12 (12 named) and lg-recall 4 (4 named), as on main, every other grid 0.
  • corpus-delta --all on the rc.21 frozen v2 corpus (34 projects, main's checker against this tree's): violations 243 -> 243, uncertifiable 470,816 -> 470,816, no row moved; needs a look: nothing. corpus-guard (ADR 1012): both builds 0 violations lost, 0 places newly uncertifiable, 0 added, no uncertifiable row moved.
  • Monotone: above. The wasm package's own tests with a FIFO in TMPDIR and without one: 80 of 80 pass in both.
  • Guards against main: count (308 keys, 37,502 rows, 481 drops) passes without --update, rises 2 measured and 2 named, regress 0 removals and 0 clean -> U, base and proved (12,671 ledger ids) 0 problems.
  • make verify: not green on this machine today, and not because of this change. First run: the builds-and-suites step failed its producer lane on one timing test, a_panic_after_the_request_stops_the_process ("producer still alive 20s after a panic"), which passes alone (the producer workspace, 310 tests and the process tests, rc 0). Second run: builds-and-suites (1,940 fuzz tests) and gates (coverage 266 misses, ownership, tsc oracle) pass, and lint-misc alone passes (kernel-oracle 25,020 cases, 0 failures), but verify-performance fails its wall-time ceiling: best first Reactive IR analysis 228,422 ns/source against 175,000 in the run, 237,526 alone (this machine passed at 169.7 and 172.0 µs earlier today). The machine reports powermode 1 and another agent was compiling. An A/B of the benchmark binary, main fdc28ba's against this build's, alternating in one sitting (three rounds, the gate in report mode, the number only): main 231,413 / 230,467 / 229,841 ns/source, this build 231,200 / 228,644 / 228,394. The same number: the ceiling failure is the machine's. A third rerun alone, while CI ran: 228,549 ns/source (1-minute load 5.3, still powermode 1). No green stamp was written, and the limit is not changed. Ruling of the coordinator: merge on CI on the strength of the same-sitting A/B (no regression from the bundle: a bundle slower than main in any A/B would stop it); verify-performance is rerun on main once the machine leaves powermode 1, and its number reported.

🤖 Generated with Claude Code

yumemi-thomas and others added 8 commits November 21, 2026 17:47
…at disposal (reads only); fixtures

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… a link on its chain (SC1001 false clean)

On main b06a4e7, a store handed by prop and held in a `const` alias
(`const held = props.s1; held.a.v = 9`) had no row of any rule at the Get
`held.a`, though rc.21 warns STRICT_READ_UNTRACKED there. Chrome: new
runtime-oracle/prop-alias-link-get is unproven-clean on main. The value-use
census (ADR 1114) took a decided SC2003 write starting at a member reached
through an alias (`write_decided`) as that member's answer; in a deeper
write the member is a Get of its own. The Get alone (`void held.a`) kept
its SC9012 row.

value_use::Census::member counts a decided write only where it is the
member's own Set (write_decided_target). A deeper write leaves the Get open:
SC9012 `value-held-in-binding` (fail closed). A violation there needs the
alias proven to be the store (ADR 1169).

Evidence:
- Fixture prop-alias-link-get: WriteChild gets the row; RootWriteChild
  (`held.items = [1]`, a Set with no link Get) is unchanged. Chrome:
  covered-uncertifiable (main unproven-clean).
- Coverage: SC9012 rises only, the same shape in local-alias-member-reads
  (`holder.state.nested.v = 1`; its direct twin is an SC1001 V) and
  store-write-literal-links (`thing.useIt.a.v = 1`). Named in rises.json.
- Probe (release, 1,837 programs: R4's N1 rows and 396 legal STORE_WRITE):
  no outcome moves.
- rediscovery.json: an `outside` row. The read grids cross no store variant
  with a prop or context source; mx-src generates
  `…-nested-assign-alias-none-s-prop` but judges it for SWOS only.
- Ownership gate 96 passed. counts.json: rises. make verify-delta: all
  passed. Not run: make verify, corpus-delta (fail-closed change),
  fuzz-smoke.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s an onCleanup callback (legal half of G3)

Coverage: async-read-under-jsx CleanupInShow and effect-cleanup-rerun-read NeverReruns SC1001 U->clean
(Chrome quiet; both READMEs already called them silent). New fixtures effect-cleanup-disposal-only,
effect-cleanup-disposal-reachable; oracle cases; guard drops rows; ADR entry and rule page.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…0491's literal proof (ADR 1172 part 2, R4 N1)

R4 group N1 (764 misuse rows, `store-write-unproven`). `props.s1.a.v = 7`,
with `<C1 s1={s1} setS1={setS1} />` and `s1` a local createStore, was
uncertifiable. ADR 0919's handed route proves `props.s1` is the store, but
the link past it needs ADR 0491's literal proof, which refused a store or a
setter used as a JSX attribute value.

ADR 0491 now accepts a hand-off to an exact local component (a JSX tag that
resolves to a function of the file; handed_component_prop):
- the store handed as an attribute whose every read in the component is a
  contained use: `props.s1`, a destructured `s1`, or an exact `const` alias
  of either, to four levels (component_prop_reads, const_alias_reads);
- the setter handed as an attribute the component never reads.
The component's props must be read only through static members, or through
a pattern of static keys with no rest, default or nested value.
handed_container_write then proves the links from the store's own binding
(handed_links_wrappable: strictNullChecks, no optional link).

Measured (release, vs main b06a4e7, Chrome cached). R4's 1,441 N1
programs plus 396 legal lg-recall STORE_WRITE:
- U -> V 428 (mx-src 208, mx-pairs 208, mx-combo 12);
- no legal row moved; no V lost; nothing needs a look.

Fixture store-write-through-prop: 4 V, 4 U (setter read and called, prop
handed on, props used whole, store also leaving the parent). Chrome covers
all eight.

Coverage: store-handed-root `Nested` goes U -> V (Chrome-confirmed on a
scratch case of it); its comment is updated.

Gates: ownership gate 96 passed. counts.json: a clean row. proof-lint: the
links-past-root filter restated. Also: open-work notes (this, and G12/ADR
1168 not built); rule page; ADR 1172 part 2. make verify-delta: all passed.
Not run: make verify, fuzz-smoke, corpus-delta.

Still U: a context, merge/omit view or spread hand-off; a component in
another file; a write in the setter's callback reached through a prop (N1's
legal in-setter rows); an SC1001 violation at an alias's Get (part 1's
SC9012 row stands, ADR 1169).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…to claude/batch-2026-11-21q

# Conflicts:
#	docs/decisions.md
#	docs/open-work.md
#	scripts/fuzz/rediscovery.json
#	scripts/guard/counts.json
#	scripts/guard/rises.json
… claude/batch-2026-11-21q

# Conflicts:
#	docs/decisions.md
#	scripts/guard/counts.json
The merges keep ours on counts.json. This is main (fdc28ba) plus each branch's
own delta against its base (1172 on 14469fa, 1167 on b06a4e7), additive, and
the branches' own drops rows (1172: 1, 1167: 2); count.mjs passes without
--update (308 keys, 37,502 rows, 481 drops).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
ADR 1167 adds a silent-capable classifier arm in read_execution_role
(execution_role.read_execution_role.returned-cleanup-disposal) and registers
neither a witness pair nor a gap, so scripts/fuzz/boundary.test.mjs ("every
silent-capable arm is separated by a pair or has a stated gap") fails on the
branch and on any join. No pair of the registry's snapshot separates the arm;
the fixtures, Chrome cases and grid probe the ADR names do, and the gap says so.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant