Skip to content

Batch 2026-11-22r: ADRs 1164, 1165 and 1178 (for-await bodies and finally after await; owner calls in a for-await body; a store handed whole to a reviewed reader) - #316

Merged
yumemi-thomas merged 33 commits into
mainfrom
claude/batch-2026-11-22r
Nov 21, 2026
Merged

yumemi-thomas merged 33 commits into
mainfrom
claude/batch-2026-11-22r

Conversation

@yumemi-thomas

Copy link
Copy Markdown
Owner

Bundle r (Mac 24c), from main a63e130 (#315). Three accepted items, each merged normally (none carries foreign history): ADR 1164 (nine commits on b600297) and ADR 1165 (stacked on 1164: its own commits and two merges of 1164; merged second, so it brings only its own work), then ADR 1178 (ten commits on b06a4e7), plus my own commits. The authors ran neither make verify nor make corpus-delta (1178 ran make verify-delta; 1164/1165's own coverage, proof gate and exec.test.mjs failed at the end on a stalled host, producer 60 s timeouts): this bundle runs all of it, the census and rest legs of the smoke included. ADR 1166 (back with its author, a false clean) and 1177 are not in.

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

  • ADR 1164 (claude/after-await-loops-1164 c55867a, base b600297), G1 of the R3 catalog: a read written directly in the body of a for await over an array literal whose first element is a number, boolean or null literal (no visible write reaches the array iterator) runs, every run past the loop's await (after_await.rs for_await_bodies_that_run); and a try head may await the global Promise.resolve(…) with literal arguments only (body_roles.rs operand_cannot_throw), so the finally after it is past the await. The read roles that answer Unreported keep the declaration-only proof; the write, owner, hand-off and reach proofs take the wider one. Author's probe (2,160 programs of mx-recall and lg-recall in the *-for-await and *-finally contexts): 188 moves, 93 legal U -> clean (SC2001, SC2006, SC2008 and SC3001 in a finally, 5 SC1002 closure reads) and 95 misuse recall-gap -> agree (89 NO_OWNER_* and SETTLED_CLEANUP_UNOWNED in a finally, 6 UNTRACKED_READ_AFTER_AWAIT in a for await body).
  • ADR 1165 (claude/owner-after-await-1165 432f82f, stacked on 1164), G4 of the R3 catalog: the owner pass also counts as after-await an operation or an onSettled call that body_roles::in_for_await_body_that_runs places (ADR 1164's condition; the function must not probe getOwner()): createMemo(async () => { for await (const _x of [1]) { createEffect(…); } }) was owned and unowned at once (owner-varies-by-caller) where Chrome reports NO_OWNER_EFFECT. Author's probe, ADR 1164's build against this branch, the same 2,160 programs: 22 misuse recall-gap -> agree (NO_OWNER_EFFECT 9, NO_OWNER_CLEANUP 9, NO_OWNER_BOUNDARY 2, SETTLED_CLEANUP_UNOWNED 2).
  • ADR 1178 (claude/g6-census-standard-1178 a96a268, base b06a4e7), G6 of the R3 catalog, steps 1 and 2: a store handed whole to a reviewed reader (Object.values, Object.entries, one-argument Array.from, Object.hasOwn on a settled store, deep, snapshot on a settled store; then [...s], { ...s }, for (const x of s), a template substitution) ends the census walk at the reader and goes on through what it answers (value_use.rs whole_store_argument, whole_store_form; two Dialect questions), so a legal program with such a read is no longer SC9012 value-in-carrier. Author's measure (2,000 of 2,884 value-in-carrier programs, Chrome cached): lg-recall 613 uncertifiable -> clean, 0 violations, mx-recall no move; Chrome silent except three body reads that were already proven SC1001.
  • Facts schema stays 108 (none of the three touches it).

Moves, read against main a63e130

  • Coverage (make coverage: nothing moved after the merge; the proof gate and the guards pass; every snapshot change of an existing fixture is a named one): ADR 1164: write-past-layered-await Named's write in the finally, SC2001 uncertifiable -> clean; new fixture after-await-for-await-finally. ADR 1165: for-await-body CleanupInBody, SC4001 owner-varies-by-caller uncertifiable -> violation; new fixture owner-after-await-for-await. ADR 1178: store-standard-library-writes (({ ...store }).v = 7, a write to a copy) loses its row (clean), store-element-flow and rc21-whole-store-reads rows relabelled or moved to the use of what a reader answers (still uncertifiable; three rises, named); new fixtures store-whole-standard-census and store-whole-standard-census-impostor.
  • Checker-only scan of the smoke's rest-leg cover (blocker-scan.mjs, no Chrome: the 45,628 programs of the cover drawn as CI draws it, minus the census grid; old 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): 45,577 identical, 51 changed. U -> V 4: mx-recall NO_OWNER_EFFECT (2), NO_OWNER_CLEANUP, NO_OWNER_CONTEXT in a finally (owner-varies-by-caller -> violation; ADR 1164/1165). U -> clean 27, all lg-recall (legal): 14 in an async-finally context (SC2001 write 5, RESOLVE_/UNTIL_IN_REACTIVE_SCOPE 5, action 2, on-flush 2: function-run-unplaced goes; ADR 1164) and 13 whole-store reads (value-in-carrier goes; ADR 1178). The others keep their class: 6 legal programs lose a function-run-unplaced row (RESOLVE_/UNTIL_IN_REACTIVE_SCOPE in a finally), 3 legal programs lose value-in-carrier rows, 5 misuse programs lose value-in-carrier rows beside a violation that stays, one misuse program loses one value-in-carrier row and keeps others, 4 relabel owner-varies-by-caller -> owner-allocation-unproven (NO_OWNER_EFFECT, SETTLED_CLEANUP_UNOWNED), one handed-member-dispatch -> handed-member-read. No misuse program loses a row without a violation kept, and the smoke's judge below agrees.
  • The oracle judge on every moved-up program (the smoke's own judge, focus.mjs per grid on the program ids with --runtime-diff all, this build's release checker): lg-recall 27 -> {clean: 27}, mx-recall 4 -> {agree: 4}: 0 false-positive, 0 unexecuted, nothing else, and the runtime differential (Happy DOM's record against Chrome's) is the same on all 31 (0 behaviour, 0 attribution-only, 0 inconclusive).
  • Monotone, old main's release checker against this build's (textbook and fixtures, 1,789 programs): 0 errors, 42 (site, code) moved up, M2 three programs with a downward move, all ADR 1178's SC9012 clean -> U at the use of what a reader answers (rc21-whole-store-reads 3, store-element-flow 2, store-whole-standard-census, its new fixture, 3), the named rises; no violation moves down.

Held out

ADR 1166 (back with its author: two census false-cleans, mac24/q-1166-false-clean.md) and 1177 wait for the coordinator's go. Nothing unaccepted is in this branch.

My commits

  • Merges of ADR 1164, ADR 1165 and ADR 1178; counts.json (e61f751): main plus each branch's own delta (1164 against b600297, 1165 against 1164's tip, 1178 against b06a4e7), additive, and the branches' own drops rows (1164: 1, 1165: 1, 1178: 2); count.mjs passes without --update (308 keys, 37,555 rows, 485 drops).
  • rises.json (d49685a): ADR 1178's three SC9012 rows that move to the use of what a reader answers (rises: 3 measured, 3 named). The boundary registry needs nothing (ADR 1178 states why: no new role arm; the registry test passes on the join).
  • Dropping a stray symlink (47ca110): ADR 1165's commit 825ce47 ("WIP ADR 1165: snapshots, guard counts, shared getOwner probe") tracks scripts/fuzz/install/node_modules -> /Users/thomas.guehenneux/EDF/Yumemi/GitHub/solid-checker/scripts/fuzz/install/node_modules, the author's machine. It is not part of the change (it dangles everywhere else); merging it replaced the harness's real, ignored install directory with it (git overwrites ignored files without a word), which made all 21 textbook programs uncertifiable until the install was provisioned again (bun scripts/fuzz/provision.mjs). The symlink is removed in this branch, so main never gets it; the author (W8) has stopped, so the 1165 branch needs no fix once this merges.

Merge notes

  • ADR 1164 and 1178: conflicts only in docs/decisions.md, .gitignore, counts.json and rediscovery.json, by my merge helper (it dropped the old ADR 1156 paragraph that a keep-both resolution brings back); no Rust conflict.
  • ADR 1165: the helper's resurrected-line rule took the merge base with HEAD (1164's tip, not a main commit) for the branch's base and dropped eleven of 1164's own ADR lines from docs/decisions.md before the commit; caught by the lost-lines count, redone as a union 3-way merge (git merge-file --union), and the helper now takes the branch's base on main. The merge itself took five minutes (git deleting the old node_modules for the symlink).
  • Text check of every branch against the result: every added line present except one docs/open-work.md note that 1165 rewrites (its version is the one in the tree) and counts.json (regenerated).

Checks (this Mac)

Everything below ran on the PR head d49685a.

  • cargo check --workspace --all-targets: clean. make coverage: 1,768 fixture projects, 40,441 findings, nothing moved. make proof-gate: ok. 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, exec.test.mjs included): pass, 1,940 fuzz tests, 0 failed, 0 skipped; the boundary registry tests pass on the join.
  • The full census smoke leg, CI's exact command: rc 0, cover 90,200 programs (sha256 8a77406c9c30a084), 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, 45,628 programs judged (mx-recall 34,294, lg-recall 6,557, mx-pairs 2,779, lg-pairs 1,119, mx-src 364, lg-src 185, mx-base 174, mx-combo 100, lg-base 30, lg-combo 26), 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, 37,555 rows, 485 drops) passes without --update, rises 3 measured and 3 named, regress 0 removals and 0 clean -> U, base and proved (12,671 ledger ids) 0 problems.
  • make verify: every step passes except the wall-time gate of verify-performance, which is the machine's. builds-and-suites (1,940 fuzz tests) and gates (coverage, ownership, tsc oracle) pass, and lint-misc alone passes (kernel-oracle 25,020 cases, 0 failures); verify-performance fails its ceiling: best first Reactive IR analysis 224,442 ns/source in the run, 224,930 alone, against 175,000. The machine has been in powermode 1 since yesterday evening (it passed at 169.7 and 172.0 µs before). A/B in one sitting, main a63e130's benchmark binary against this build's, alternating three rounds (gate in report mode, the number only): main 225,227 / 224,179 / 223,566 ns/source, this build 223,249 / 223,470 / 223,611: the same number, no regression. No green stamp was written and the limit is not changed; per the coordinator's ruling (as for Batch 2026-11-21q: ADRs 1172 and 1167 (a store written through a prop; a never-rerunning effect's returned cleanup) #315) this merges on CI with that A/B, and verify-performance is rerun on main once the machine leaves powermode 1.

🤖 Generated with Claude Code

yumemi-thomas and others added 30 commits November 21, 2026 13:23
…Promise.resolve(), are past the await

Slice report (G1 S half; the WIP commits on this branch are this slice).
- after_await.rs: a read written directly in a `for await` body over an
  array literal led by a number/boolean/null literal (array iterator intact)
  is no longer marked SuspensionUnplaced: SC1002 violation.
- body_roles.rs: the try-head proof also takes the global
  `Promise.resolve(literals)` (exact resolution to
  PromiseConstructor.resolve, Promise intrinsics intact); `new` operands are
  refused for both forms. The read roles' Unreported arms keep the
  declaration-only proof, directly and through a climb (ADR 1112's flight
  question).
- Fixture after-await-for-await-finally (+ runtime-oracle twin);
  write-past-layered-await Named's finally write U -> clean (guard drops row).

Checks: Chrome on the fixture (V at the three reads Chrome reports, nothing
at the clean writes, every U that runs covered). Probe (release, VM, main
b600297 vs this branch, 2,160 mx/lg-recall programs in the for-await and
finally contexts): 93 legal U -> clean, 95 misuse recall-gap -> agree,
nothing else, nothing needs a look. make coverage, ownership gate: pass.
verify-delta: every check passes; fuzz-tests failed twice under load only on
exec.test.mjs's 0.8 s readiness (passes alone and in a later full lane run,
1940/1940); proof-gate failed once in a sequential run and passed when rerun
alone.
Not done: the legal half of G1 (reads in an async function that only a
handler or the body calls: no "no flight" reason, ADR 0973 ruling 5).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Slice report (G4 S half, stacked on ADR 1164; the WIP commits on this branch
are this slice).
- owner_emission.rs: an effect/cleanup/context read/boundary call, or an
  onSettled call, that body_roles::in_for_await_body_that_runs places (the
  ADR 1164 literal condition) counts as after-await, unless the function
  probes getOwner() (shared probe with the climb).
- Fixture owner-after-await-for-await (+ runtime-oracle twin);
  for-await-body CleanupInBody U -> V (guard drops row).

Checks: Chrome on the fixture (NO_OWNER_EFFECT/CLEANUP covered at the new
violations; SETTLED_CLEANUP_UNOWNED raised unattributed, so SettledInBody
reads static-only there; the one-site grid programs judge it agree). Probe
(release, VM, ADR 1164's build vs this branch, same 2,160 programs): 22
misuse recall-gap -> agree, nothing else, nothing needs a look. Coverage
--update and the ownership gate passed before the final commit's pure
refactor (one getOwner closure shared). verify-delta: fmt, clippy, guard
count, read/proof lints, ir-lib, wasm-compile, diagnostics_process,
write_walk_strategies, repro-causes, ownership gate, tsc-oracle gate,
obligation audit, witnesses, repro-budget, ir-edit-fuzz pass; rewrite.test
passes alone. NOT passed on the host at the end: coverage and proof-gate
runs failed on producer handshake timeouts (60 s) under machine load
(single-fixture runs take 1 s; the filesystem stalled), and exec.test's
0.8 s readiness. These need a rerun on a quiet machine.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gh what it answers (R3 catalog G6, steps 1 and 2)

The value-use census (value_use.rs) now ends a store root's walk at Object.values, Object.entries, one-argument Array.from
and deep where the whole-store read row (ADRs 0620, 0621) stands at the call; at Object.hasOwn and snapshot on a store
settled from its creation (Dialect::creates_settled_store_from_value; hasOwn also where ADR 0635's key-probe row stands);
and at [...s], { ...s }, a for…of with identifier bindings and a template substitution where ADR 0621's row stands. It
walks on through the result: arrays and object copies as a root holder, so an element Get keeps its row; deep/snapshot
results as the root. Standard callees must be exact, not withheld (ADR 0840) and not replaced (ADR 1129), and the same
premise is asked of the iterator's and the coercion's members. New Dialect question Dialect::whole_store_argument
(deep = ReadsEveryKey, snapshot = ReadsRawBacking). Proof id value_use.whole_store_argument (ablatable).

Measured (release, b06a4e7 vs this, sc VM, Chrome cached): 2,000 store-* programs picked from a 3,996-program pool by
the base's value-in-carrier rows. lg-recall 1,112: 613 uncertifiable -> clean, 479 stay U, 20 typed away, 0 violations.
mx-recall 888: no move (139 agree, 749 recall-gap). R5's ~668 + ~348 grid rows is an estimate; only the sample is measured.

Chrome: runtime-oracle/store-whole-standard-census (release build) warns STRICT_READ_UNTRACKED only at the three BodyForms
reads, each a proven SC1001; silent elsewhere; no unproven clean.

Checks: make coverage (5 projects moved, reviewed: store-standard-library-writes `({ ...store }).v = 7` clean as its README
says; store-element-flow and rc21-whole-store-reads rows relabelled/moved, still U; two new fixtures), coverage-update,
ownership gate 96 passed, guard count (two drops rows), proof-lint (allowlist reason for the fall-back arms),
make verify-delta BASE=b06a4e71a in the sc VM clone at c5ea3d8: all 28 checks passed (fuzz tests 1,940 passed,
0 skipped). Not run: make verify (the integrator's), corpus-delta (no veto change), a full-grid sweep.

Remaining U: step 3 `[s].map(cb)` (not built); results read further, returned from a primitive callback or handed on;
props roots; helper parameters; snapshot/hasOwn on a derive, optimistic store, projection or non-literal store; call
spreads; pattern and for-await loop bindings; programs that replace a standard member the form runs.
…nto claude/batch-2026-11-22r

# Conflicts:
#	docs/decisions.md
#	scripts/guard/counts.json
…nto claude/batch-2026-11-22r

# Conflicts:
#	scripts/guard/counts.json
The merges keep ours on counts.json. This is main (a63e130) plus each branch's
own delta (1164 against its base b600297, 1165, stacked on 1164, against 1164's
tip), additive, and the branches' own drops rows (1164: 1, 1165: 1); count.mjs
passes without --update (308 keys, 37,525 rows, 483 drops).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…acks

scripts/fuzz/install/node_modules is a tracked symlink to
/Users/thomas.guehenneux/EDF/Yumemi/GitHub/solid-checker/scripts/fuzz/install/node_modules,
the author's machine, added by "WIP ADR 1165: snapshots, guard counts, shared
getOwner probe" (825ce47). It is not part of the change: on any other machine it
dangles, and merging it replaced the fuzz harness's real (ignored) install directory
with it, which made every textbook program uncertifiable (SC9014 at solid-js's
package.json) until the install was provisioned again. node_modules stays an
ignored, provisioned directory (bun scripts/fuzz/provision.mjs).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
yumemi-thomas and others added 3 commits November 22, 2026 01:52
…into claude/batch-2026-11-22r

# Conflicts:
#	docs/decisions.md
#	scripts/fuzz/rediscovery.json
#	scripts/guard/counts.json
The merges keep ours on counts.json. This is main (a63e130) plus each branch's
own delta (1164 against b600297, 1165 against 1164's tip, 1178 against its base
b06a4e7), additive, and the branches' own drops rows (1164: 1, 1165: 1, 1178: 2);
count.mjs passes without --update (308 keys, 37,555 rows, 485 drops).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ader answers

rc21-whole-store-reads SC9012 value-from-primitive-callback 0 -> 3 and
store-element-flow SC9012 value-held-in-binding 0 -> 2, value-root-in-helper
8 -> 9: the census now follows a store handed whole to a reviewed reader through
what it answers, so the row that stood at the reference stands at the use, still
uncertifiable (the matching falls are the branch's drops rows in counts.json).
rises: 3 measured, 3 named.

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