Repository navigation
Batch 2026-11-22s: ADRs 1174, 1177, 1169 and 1155 r4 (refresh of a prop; a context read in a handler; a store link held in a binding; a forwarded leaf operation) - #317
Merged
Conversation
…kes no provider claim
Census false positive (mx-context-not-found-dny-flush-{load,popover,unmount},
open-defects, Mac 23): useContext in an onLoad/onBeforeToggle/onBlur handler a
later flush reaches got SC4002 (ContextNotFoundError); Chrome throws
NO_OWNER_CONTEXT first (useContext checks the owner first). ADR 0577's
owned-run-first premise holds for a body call beside a handler, not for a
handler literal whose deliveries are owned on one stack and unowned on
another. competing_checks: a reported context-read requirement whose owner
varies by caller or may be null, at a read in a JSX handler, makes the
provider claim uncertifiable (owner-check-throws-first).
Checks: fixture context-read-owner-varies (Varies V -> U; Body and the ADR
0577 BodyAndHandler stay V); Chrome case (NO_OWNER_CONTEXT at Varies,
CONTEXT_NOT_FOUND at the other two). Fuzz, release 14469fa vs this build,
CONTEXT_NOT_FOUND + NO_OWNER_CONTEXT over mx-base, mx-combo, lg-base,
lg-combo, mx-census (6,440 programs): 3 false positives -> not-fired; 34
census agree -> recall-gap (accepted cost, owed order proof in open-work);
nothing else. Coverage: new fixture only. open-defects rows and ratchet
updated; rediscovery entry added. Ownership gate 96 passed, verify-delta
passed, rises.mjs 0. Deferred: corpus-delta (running), fuzz-smoke, make
verify.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eshable source, placed by its role Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…every caller passes (R4 N3) Probe vs main fdc28ba (mx-src memo/compute cells with prop and context sources, 2,394; lg-src legal, 252), every finding of every code: 180 programs, each one SC2001 U->V; nothing else moved. Chrome rc.21 on the fixture: covered-violation at the three V sites, quiet at the handler. Fixture refresh-prop-target, ADR entry, rule page, open-work line. No new role arm (boundary test passes). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…back is reported where written (ADR 1155 r4)
Bundle p's follow-up: forwarded-async-read SyncApp's `sync: true` createEffect,
in `hand(() => …)` with `hand(fn) { createTrackedEffect(() => { fn(); }); }`,
was SC7002 uncertifiable (creation-throws-first) and SC3001
leaf-operation-unaccounted, while Chrome throws PRIMITIVE_IN_FORBIDDEN_SCOPE
there, which SC3001 (leaf-owner-forbidden-call) reports.
The leaf walk now also reports a call the leaf refuses written directly in the
body of a literal forwarded_only_into_leaf proves
(callee_roles::forwarded_leaf_call), when neither the literal nor the leaf
callback is async or a generator and the wrapper's call of its parameter is in
the callback's body. SC7002 stays uncertifiable there.
Fixture reactive-ir/forwarded-leaf-operation (Chrome
runtime-oracle/forwarded-leaf-operation): SyncApp, MemoApp, CleanupApp,
FlushApp, NamespaceApp, MemberApp V (each covered-violation); SignalApp,
ShadowedApp, MemoComputeApp no row; NestedApp, AsyncLiteralApp no violation
(Chrome silent); BothApp (arguments[0]()), SettledApp (onSettled), RelayApp
(handed on) U (Chrome throws: true defects left uncertifiable).
Coverage: forwarded-async-read SyncApp SC3001 U->V, nothing else.
counts.json: drops row (proof callee_roles.rs::forwarded_leaf_call).
Probe (release, Chrome cache), against 1155 r3: census barm-fwdleaf contexts
(126) recall-gap 31->24, agree 20->27; the forbidden-scope codes on every grid
(7,444, legal included): no other move except two lg-combo U->clean that are
clean with this pass switched off too (main's own move since 1155 r3).
No V->U, no legal V.
Checks: make coverage + update, ownership gate (96 passed), guard count,
make verify-delta. Not run: make verify, corpus-delta, fuzz-smoke. No
rediscovery row: a recall gain, no defect (the census leg runs these cells).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nto claude/batch-2026-11-22s # Conflicts: # docs/open-work.md # scripts/guard/counts.json
…to claude/batch-2026-11-22s # Conflicts: # docs/decisions.md # scripts/fuzz/rediscovery.json # scripts/guard/counts.json
…' into claude/batch-2026-11-22s # Conflicts: # docs/decisions.md # scripts/guard/counts.json
…01 false clean) On main 14469fa and ADR 1172's tip, `const held = s1.a; void held.v` had no row of any rule at the Get `held.v`, though rc.21 warns STRICT_READ_UNTRACKED there. Chrome: new runtime-oracle store-link-alias-get is unproven-clean on both. The value-use census answered a store root's own member chain (Mode::Root) where the chain ended, and never walked a binding that holds the chain's link. value_use::Census::member: the root's own static chain, held whole in an identifier binding, is walked on as the root. A Get through the alias with no read row stands open: SC9012 value-held-in-binding (fail closed). Part 2 proves an exact alias as the store's link. Evidence: - Fixture store-link-alias-get: two alias Gets go from no row to U; the member twin stays an SC1001 V. Chrome covers them. - Coverage: SC9012 rises only, in 11 existing fixtures (after-await-*, forwarded-prop-backing, props-callers, returned-store-nested-read, store-copy-element-write, …). Every move read; no finding removed; all named in rises.json. - Probe (release, 1,998 programs: mx-recall nested-held/sw-let/ nested-assign/nested-push, lg-recall store-*): no outcome worse. The 55 U -> clean are ADR 1162's, against R3's older base. - rediscovery.json: an `outside` row (no read variant holds a link). The generator gaps (this, and prop/context crossing) are noted in open-work, deferred. - Ownership gate 96 passed. counts.json: rises. make verify-delta: all passed. Not run: make verify, corpus-delta, fuzz-smoke. Stacked on ADR 1172 (claude/prop-store-write-1172), which is not on main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 0e4e1ee)
…as of a proven literal store is V A `const held = s1.a` over a function-local createStore of a closed literal whose setter is unreferenced (ADR 0491's literal proof) is the store's link: - SC2003: a handed or local write through `held` past the root is a violation (container_root HandedRoot::Store gains `below`; handed_links_wrappable follows the alias via link_alias_literal; a contained alias reference keeps the store's proof via link_alias_contained). - SC1001: `held.v` in a strict window reads as `s1.a.v` (local_access feeds a synthetic member for the alias; the message names `held.v` as written). - A let, array, handed-on alias or a referenced setter stays uncertifiable; a prop's alias (`props.s1.a`) stays SC9012 because its direct twin is SC1001 U. Fixtures: reactive-ir/store-link-alias-write (IifeWrite, TimerWrite, DeepWrite, AliasRead V; SetterReferenced, HandedOn, LetAlias, ArrayAlias U), Chrome runtime-oracle/store-link-alias-write covered; READMEs for store-link-alias-get and prop-alias-link-get. Coverage moves read: store-link-alias-get 11/26 U->V (Chrome confirmed); flush-write-path Alias SC2003 U->V (Chrome covered-violation); relabels U->U in after-await-read-first-writes and store-sub-value-flow. No V->U, no legal V. count.mjs: relabeledTo accepts a list (a split relabel). proof-lint: default-on-missing in shared_reactivity.rs fell 13 -> 12 (baseline shrunk). Probe (release, mx-recall/lg-recall ids vs part 1): mx-recall U->V 51, lg-recall unchanged, no V->U, no legal V. Checks: make coverage + update, ownership gate (96 passed), make verify-delta (every selected check passed). Not run: make verify, corpus-delta, fuzz-smoke. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 6247fcb)
ADR 1155 r4 (27b0df6) appends its paragraph at the end of ADR 1155's section, and ADR 1174 appends its whole entry at the same place; the union merge put 1174's entry first, which left "Round 4 (Mac 27; bundle p's follow-up)" under ADR 1174. Moved after round 3's evidence line, one line, nothing else changed. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The merges keep ours on counts.json. This is main (960bfa1) plus each branch's own delta (1174 against fdc28ba, 1177 against 14469fa, 1155 r4 against fdc28ba, 1169's own two commits against 1172's tip 286b374), additive, and the branches' own drops rows (1155 r4: 1, 1169: 3); count.mjs passes without --update (309 keys, 37,677 rows, 489 drops). Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
corpus-guard on the bundle (main 960bfa1 + ADR 1174, 1177, 1155 r4, 1169): both builds 0 violations lost, 0 added, 6 places newly uncertifiable (SC9012 in compass-ui 1, inferay 1, kui feature-clusters 1, kui shell 3). They are named here, std and dev, citing ADR 1169 part 1. Checked first: the bundle adds 805 SC9012 rows corpus-wide (value-held-in- binding 372, value-in-carrier 182, value-to-unresolved-callee 104, value-root-in-helper 96, value-through-logical 24, value-use-unclassified 15, value-from-primitive-callback 8, value-through-new 4), none replaces a violation, none sits where the old build's proof log names a value_use proof. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The full census leg of the smoke on bundle s (main 960bfa1 + ADR 1174, 1177, 1155 r4, 1169) exits 1 on ADR 1177's accepted cost: 33 census programs move agree -> recall-gap ("worse" against the manifest) and the same 33 are proved-ledger programs lost with no changes.json row. The author ran no fuzz smoke, so neither file had it. Coordinator ruling: zero false positives outranks the 33 withheld true proofs (the shape of the ADR 1125 and 1139 rows); record. - scripts/fuzz/smoke/mx-census.json: the expectation of the 36 programs whose behaviour the bundle moves and the manifest holds: 33 agree -> recall-gap (ADR 1177) and the 3 false positives dny-flush-{load,popover,unmount} -> not-fired. Not `smoke.mjs record --whole`: the manifest is the whole grid (51,203 ids), the cover run holds 44,572, so that would drop ids and re-record main's 209 older better-or-changed expectations. Every other id and expectation is kept (decoded diff: exactly these 36; 51,203 ids). - scripts/fuzz/proved/changes.json: one fix row, ADR 1177, the 33 ids, witness dny-flush-load (Chrome NO_OWNER_CONTEXT where main reported CONTEXT_NOT_FOUND), collateral 33. - scripts/fuzz/proved/owed.json: ADR 1177 owes 33 (the order proof of the open-work line). Census run, main against bundle: 34 agree -> recall-gap (33 here, plus dny-change-checkbox-retyped, which is in neither the manifest nor the ledger), 3 false-positive -> not-fired, 5 recall-gap -> agree (ADR 1155 r4, barm-fwdleaf-tracked; not in the manifest), 0 false-positive/false-clean. After the record: moved from the manifest 0 worse, 209 better or changed (main's own); proved verdicts 37 lost, 37 named, 719 newly proved; scripts/guard/proved.mjs 0 problems. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Bundle s (Mac 24c), from main 960bfa1 (#316). Four accepted items: ADR 1174 (three commits on fdc28ba), ADR 1177 (one commit on 14469fa), ADR 1155 r4 (one commit on fdc28ba) and ADR 1169 (its own two commits, cherry-picked: the branch is stacked on ADR 1172's pre-squash commits, which main has as #315), plus my own commits. None of the four ran
make verify,make corpus-deltaor the fuzz smoke; this bundle runs all of it, the census and rest legs of the smoke included, and judges every moved-up program against Chrome. ADR 1166 r2 was in the first build of this bundle and is out (below). ADR 1173, 1176 and 1179 wait for the next bundle.In the bundle (each accepted by the coordinator on evidence)
refresh(props.name)in a memo, effect compute or render compute, the argument a prop, was SC2001function-run-unplaced(630 of 630 refresh programs of mx-src's memo-compute cells).static_api::props_refresh_source: the target isprops.nameon a component's own props parameter that is current with only views attached (or a never-written destructured slot), the caller census enumerates every caller, each passing that one exact refreshable source; the write is then that source's and takes the call's role. Author's probe: 180 programs of mx-src change, each one SC2001 uncertifiable -> violation; lg-src's 252 legal programs unchanged; Chrome: covered violation at the three.mx-context-not-found-dny-flush-{load,popover,unmount}: auseContextof a default-less context in a JSX handler (onLoad,onBeforeToggle,onBlur) that a later flush reaches got SC4002ContextNotFoundErrorwhile Chrome throws NO_OWNER_CONTEXT first (useContextchecks the owner first).competing_checks: a reported context-read requirement whose owner varies by caller or may be null, at a read in a JSX handler, makes the provider claim uncertifiable (owner-check-throws-first). The three false positives become not-fired (open-defects emptied of them, ratchet 3 -> 0); the author names a cost, 34 census programs agree -> recall-gap, which the smoke fails until it is recorded: recorded here, see "ADR 1177's cost" below.const held = s1.a; void held.vhad no row of any rule at the Get (Chrome warns STRICT_READ_UNTRACKED);value_use::Census::membernow walks the held chain as the root and an unproven Get stands open as SC9012value-held-in-binding(14 rises named). Part 2: a write or read through an exactconstlink alias of a proven literal store is a violation. Author's probe: mx-recall 51 U -> V, lg-recall none.scripts/guard/count.mjsgains a list forrelabeledTo.callee_roles::forwarded_leaf_call): SyncApp's SC3001 is a violation (SC7002 stays uncertifiable). Author's evidence: census contextsbarm-fwdleaf-*recall-gap 31 -> 24, agree 20 -> 27; the forbidden-scope codes on every grid (7,444 programs): no V -> U, no legal V.ADR 1177's cost, recorded (coordinator ruling: zero false positives outranks the withheld true proofs)
The full census leg of the smoke (CI's command) exits 1 on ADR 1177's own named cost: 33 census programs move agree -> recall-gap ("worse" against the manifest), and the same 33 are proved-ledger programs lost with no
changes.jsonrow. The author ran no fuzz smoke, so neither file held it. Recorded in one commit (the shape of the ADR 1125 and 1139 rows):dny-flush-{load,popover,unmount}: false-positive -> not-fired, no findings now) against 33 true violations withheld (agree -> recall-gap; all CONTEXT_NOT_FOUND programs Chrome confirms:dny-*30,evt-*2,pos-handler-owner1), owed to the order proof (a run at mount, before the handler can be delivered, is the first), the author'sdocs/open-work.mdline. A 34th program,mx-context-not-found-dny-change-checkbox-retyped-x-direct-none, moves alike and is in neither the manifest nor the ledger (the author's 34).scripts/fuzz/proved/changes.json: one fix row, ADR 1177, the 33 ids, witnessdny-flush-load(Chrome NO_OWNER_CONTEXT where main reported CONTEXT_NOT_FOUND), collateral 33.scripts/fuzz/proved/owed.json: ADR 1177 owes 33.scripts/fuzz/smoke/mx-census.json: the expectation of exactly the 36 programs it holds that the bundle moves (33 agree -> recall-gap, 3 false-positive -> not-fired); decoded before and after, nothing else changed and all 51,203 ids are kept. It is notsmoke.mjs record --whole: the manifest is the whole grid and the cover run holds 44,572 of its ids, so that would drop ids and re-record main's own 209 older better-or-changed expectations, which this leaves as they are.barm-fwdleaf-tracked, not in the manifest), nothing else; the census has 0 false-positive, false-clean and wrong-rule rows. With the record, on the same run: moved from the manifest 0 worse (209 better or changed, main's own), proved verdicts 37 lost, 37 named, 719 newly proved;scripts/guard/proved.mjs0 problems.Moves, read against main 960bfa1
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 1174: new fixturerefresh-prop-targetonly. ADR 1177: new fixturecontext-read-owner-variesonly. ADR 1169:flush-write-pathAliasSC2003 uncertifiable -> violation;after-await-read-first-writesandstore-sub-value-flowrows relabelled (still uncertifiable); SC9012 rows added in eleven fixtures (the 14 named rises); new fixturesstore-link-alias-getandstore-link-alias-write. ADR 1155 r4:forwarded-async-readSyncApp SC3001 uncertifiable -> violation; new fixtureforwarded-leaf-operation.--runtime-diff all) on the 4 moved up: 4 agree, 0 false positive, 0 unexecuted, 0 differences.focus.mjs --smoke --smoke-leg census|rest --cover all --cover-seed 1 --sample 0): rest leg exit 0, nothing needs a look (mx-recall 34,294 programs, mx-pairs, mx-src, mx-combo, mx-base, lg-*: 0 behaviour or attribution differences with the oracle, 0 proved verdicts lost beyond the named); census leg: the moves above.context-read-owner-variesSC4002 V -> U (ADR 1177's fixture);forwarded-leaf-operationSC3001 clean -> V is ADR 1155 r4's.leaf-operation-unaccounted7 in 3 projects, SC6002effect-creation-unaccounted7 in 3, SC1002read-after-await-unaccounted5 in 2, SC9012value-to-unresolved-callee1). The 20 places that went away, attributed (the coordinator's condition for merging): 26 removed rows in four projects (kui shell 13, app-game 8, kui feature-clusters 4, feature-consumers 1). 15 are relabels: the rule keeps a row at an overlapping span with another blocker (SC9012value-to-unresolved-calleetovalue-through-logicaland the like; SC3001 and SC6002 at BrokerCard.tsx:177 and :185). 11 have no row of their rule left at the site: SC1002read-after-await-unaccountedx5 (app-game popover.tsx:141; kui TopicsRoute.tsx:1058, 1059 twice, 1115), SC3001leaf-operation-unaccountedx3 and SC6002effect-creation-unaccountedx3 (app-game createDragSensor.tsx:554; kui GroupDetail.tsx:281 and :282). None is uncertifiable -> clean and none is a silent drop. At all 11 the proof log's universe site now ends incensus(credited,reactive-dispatch-unresolved:value-to-unresolved-calleeorvalue-in-carrier; main's ended in the backstop's home row or inrule:action-called-in-owned-scope:u), and an SC9012 uncertifiable finding stands at the same span (popover, createDragSensor, GroupDetail) or at the enclosing call (TopicsRoute): the "unaccounted" rows are the backstop for sites nothing else ends, and a census ending is a typed ending. Every site stays uncertifiable. No proof arm or silence is new at those sites (0 candidate names in the log), so the ablation switch has nothing to turn off; the controlled build says it instead: the bundle with only ADR 1169 part 1's 21-linevalue_use.rshunk (9e0e637, the only bundle commit that touches that file) reverse-applied has 0 removed rows against main in shell and app-game, so all 21 rows of those two projects are that hunk's: the held chain's census rows now cover those Gets.scripts/guard/corpus-changes.json(ADR 1169 part 1; I checked first that none replaces a violation, a proved clean or a place whose proof log names avalue_useproof).Held out: ADR 1166 r2
claude/named-callback-1166 72fe8ad was merged last so that any finding of its own would drop it, and one showed:
corpus-delta --allsaid "needs a look (2)", kuifeature-consumersandshell, SC1001 uncertifiable -> violation, four rows, all atfeature-consumers/src/GroupRoute.tsx:80(const refresh = ...inside the apply callback ofcreateEffect(compute, apply),refresh(false)right after the declarator, and a timer re-call from inside itself). Main hasSC1001 uncertifiable function-run-unplacedx2 there; with 1166 r2,violationx2; with itsreporting-runarm ablated the uncertifiable rows are back, so it alone makes them violations. The shape (a named callback run directly in an effect's apply and again from a timer) is in none of its evidence and the corpus cannot run in Chrome: Mac 26 has the repro inmac24/s-1166-corpus.md. Everything else of its run held (the Chrome judge on all 3,960 programs it moved up: agree or clean on every one, 0 differences). Per the coordinator's rule it is dropped and the rest ships. ADR 1173, 1176 and 1179 (the next bundle) and everything unaccepted wait for the coordinator's go.My commits
count.mjspasses without--update(309 keys, 37,677 rows, 489 drops); rises: 17 measured, 17 named.docs/open-work.md(two notes each side rewrote) andscripts/guard/rises.json(rows appended by both sides, by a JSON-aware resolver).changes.jsonandowed.jsonrows above, one commit.Merge notes
docs/open-work.md(1174, 1169's second pick) andscripts/guard/rises.json(1169's first pick); the rest by my helper and resolvers.Checks (this Mac)
On the final tree 5ab8d70 unless said (the census record is data files only; the engine is the tree the gates before it ran on):
cargo check,make coverage(nothing moved),make proof-gate,scripts/guard/proved.mjs(0 problems),count.mjswithout--update, rises 17 of 17 named,make textbook-check21 of 21: pass.sh -n scripts/*.sh,vitest scripts/*.test.mjs,bun scripts/fuzz/test-lane.mjs): pass, 1,940 fuzz tests.make verify: every step passes (fmt, typefacts, builds-and-suites with 1,940 fuzz tests, gates: coverage, ownership-gate, tsc-oracle-gate, obligation-audit, witnesses, proof-gate, wasm-parity, ir-edit-fuzz, bun-test, scripts-test) exceptverify-performance: 230,304 ns/source against the 175,000 ceiling at a 1-minute load of 14.6, rerun alone 224,090 (load 6.6), the machine inpowermode 1(pmset) as on Batch 2026-11-21q: ADRs 1172 and 1167 (a store written through a prop; a never-rerunning effect's returned cleanup) #315 and 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. Same-sitting A/B (the coordinator's rule:verify-performancein report mode, main 960bfa1's benchmark binary against the bundle's, alternating three rounds, powermode 1, load 3 to 10), ns/source: main 224,009 / 224,032 / 224,032, bundle 224,920 / 224,841 / 224,505, so the bundle is 0.2 to 0.4% slower (mean +731 ns, +0.33%). Accepted by the coordinator. After the merge, at powermode 0,verify-performanceon main 01d5fdc passes the enforced gate at 166,217 ns/source (limit 175,000, headroom about 8.8k, one cold sample at load 7.5).lint-miscalone: pass. No green stamp from this run.docs/open-work.md, and ADR 1177's owed order proof is now a ledger row).🤖 Generated with Claude Code