RQ-70-NPA (#1331): static-data f32 relocation, the four gate-potency lanes, two deferrals measured, and the v0.70 steps-1-2 filing - #1340
Merged
Merged
Conversation
…native-pointer ABI — cpetig's blocked exports go 4 of 6 to 2 REPRODUCED on the reporter's own command line first, and their issue excerpt UNDERCOUNTS exactly as #1318's did: reported 2 functions, measured 6 of 22 (4 of 6 EXPORTS), rc=1 via #952, NO object — 4x f32.load and 3x f32.store "from/to the static-data region under the native-pointer ABI". AFTER: 2 of 22 functions, 2 of 6 exports. `rate#tick` and `attitude#tick` compile; zero f32 static-data declines remain. The machinery existed. `emit_wasm_data_addr` relocates a base to `__synth_wasm_data + offset`; the i32 sub-word arms got it in #744 and the i64 arms in #746. The float arms declined loudly instead — and correctly: the raw `[R11 + addr + #offset]` path BAKES the linmem offset as an un-relocated MOVW/MOVT immediate, invisible to both the #678 `--shadow-stack-size` down-shift and the post-link in-range oracle because each walks RELOCATIONS. That is the #739 silent OOB, so this ships the relocation, not the bake. No float address machinery was needed: both arms already route the 4 bytes through a core register and bit-cast (VMOV), so only HOW THE ADDRESS IS FORMED changed and the loaded/stored bytes are identical by construction. NOT CLAIMED CLOSED. `position#tick` and `ekf#estimate` now reach a DIFFERENT ceiling the f32 decline was hiding: `LdrSym literal pool out of range (#345): imm12=11744 / 4344 > 4095`. Did the relocation's own pool entries cause it? The magnitude answers — imm12 is the PC-to-pool distance, so 11744 is a ~11 KB function, and each relocated access adds at most one 4-byte pool word; even 50 added entries leaves ~11.5 KB, still ~3x over the limit. A pre-existing function-size wall, unmasked rather than introduced. RED-FIRST: scripts/repro/npa_static_f32_1331.wat, a 20-line reduction carrying one dynamic-index f32.load and one f32.store above the SP global's initializer. Pre-fix 2 of 2 skipped with exactly the reporter's two decline strings; post-fix an 813-byte object. Not on EXPECTED_DECLINES, so the wired sweep reds on a regression. The CONST static-data address declines (#237) are deliberately untouched: a different address mode, measured to fire on none of the seven. GATES: fmt, clippy -D warnings, ARM corpus sweep PASS (174/194 compiled, 3428/3428 vectors, 0 mismatches). Refs #1331 Refs #1318 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…s" turns out to name THREE populations v0.69 added only PAGESIZE's two .wat to scripts/repro; cpetig's fused.wasm and opt.wasm were never committed, so FALCON's 826 identical / 209 identically failing / 0 changed and its 1-of-21 / 7-of-17 counts cannot be reproduced from this repository by anyone — and FALCON was that release's only must. ROUTE B, TAKEN TWICE. Both of v0.70's downstream blockers now ship a committed reduction that reproduces the reporter's decline: spill_slot_alias_1321.wat (54 lines, with RQ-70-ALIAS) and npa_static_f32_1331.wat (20 lines, here). Neither is on EXPECTED_DECLINES, so the wired sweep reds if either regresses — verified by running it against the pre-fix binaries. RESIDUAL, NAMED: a reduction reproduces the DECLINE, not the module. FALCON's 826/209/0 stay non-re-derivable, and this does not claim otherwise. THE SECOND FINDING. A census of every committed script walking scripts/repro/ finds THREE populations, not the two the plan noted: *.wat only (7 scripts, 193 modules), *.wat + *.wasm (5 scripts, 209), *.wasm only (1). That is how FALCON's '1035 = 207 x 5' and SUBTRACT's '965 = 193 x 5' were both 'the corpus' in one release. And 1035 is ALREADY STALE — re-derived today the populations are 193 and 209, because v0.69 itself added two .wat. NOT unified, deliberately: the globs differ for real reasons (byte-triage needs two binaries per module; a .wasm has no second source), and collapsing them would move numbers across a dozen pinned gates for no correctness gain. The defect is the unnamed attribution in PROSE, so the deliverable is the census. Refs #1318 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…elease artifact — recursive glob, the keys artifacts actually cite under, and two namespaces instead of one The glob was non-recursive, so 105 of 135 artifact files were invisible, including ALL 10 of v0.69's and ALL 8 of v0.70's. It read only `run:` keys, which no release artifact has ever had. And it knew one citation form while release artifacts use two: `--test <target>` names an integration-test FILE, `--lib`/`--` are substrings over test PATHS — measured, 4 of v0.69's 5 citations are targets, so widening the glob ALONE would have produced four FALSE FAILURES on valid citations. RED-FIRST: planting `--test this_test_does_not_exist_9999` into implemented RQ-69-ELFNAMES left the gate at rc=0 printing 'every cited test filter resolves'. After the fix the same plant reds, naming artifact, status and missing target. The widening then exposed a real scan bug: module paths were built only from `mod` declarations, never from the FILE, so `--lib dwarf_line::` read as dangling though dwarf_line.rs holds 3 tests and lib.rs declares the module. Fixed; test names 5233 -> 7224. The header promises approximation failures are loud, and this one was. Keys are limited to run/done-when/verified-by, NOT description: RQ-69-PROSEGATE refuted prose-scanning one release ago — it cannot tell a CLAIM from a MENTION. AFTER: 135 files, 33 citations (15 target, 18 filter), 0 false claims. Refs #1333 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
… delivery convention — teach it RQ-NN, and put the floor where it can say why Re-derived per window with the script's own regex: delivery-shaped commits went 4, 8, 8, 11, 2, 0, 0 while RQ-style subjects went 3, 3, 2, 0, 3, 8, 8. The convention became 'RQ-NN-NAME (#issue): ...' around v0.67 and R10's window loop only matched 'type(scope): ', so the completeness floor that exists to notice work landing with every artifact silent saw ZERO commits for two releases. The recognizer already existed — ARTIFACT_ID is what R4 uses, which is why 88 delivery commits match overall. It was simply never applied in the window loop. CI could not have caught it: the grep is '[0-9]+ delivery-shaped', and [0-9]+ matches 0 — #1243's shape, a floor regex that does not floor. So the floor moved into the script, conditional on window size (6) because early in a cycle a window legitimately holds only a plan commit. PROVEN BOTH WAYS: replaying v0.68.0..v0.69.0 now reports 8 delivery-shaped, 8 attributed (was 0); rewriting those subjects RQ-69- -> DELIVERED-69-, simulating the next convention change, FIRES the new floor. Refs #1334 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…ATIONS and printed them as evaluation Manual share per release: v0.61 11/15 ... v0.66 8/8, v0.67 7/7, v0.68 9/9, v0.69 9/9 — four releases at 100%, all printing the same 'N/N done-when' line as a release with everything machine-checked. A `manual:` signature is present and evaluated by nothing; status_evidence's R3 fires only on contains:/file:. NOT a ban. Reading all nine v0.69 predicates: each is a COMPOUND acceptance statement (red-first AND a class sweep AND a named ceiling AND an execution differential) that no single mechanical signature can carry. But most hide a mechanical conjunct written in prose and checked by nothing — ELFNAMES, PAGESIZE, PMPLIB and VFPUNKNOWN each name a real test, SUBTRACT's line IS a claims.yaml ratchet. Five citations no gate read, which is RQ-70-CITEGAP's finding from the other side. The evidence existed; manual: was where it went to not be checked. So the deliverable is honest output. On v0.69.0 the gate now prints '9/9 done-when DECLARED — 0 mechanically evaluable (contains:/file:), 9 manual: (declared only; R3 cannot fire on these)'. Verdict unchanged: still CONFORMS. Pinned by two new unit tests that ci.yml runs. Refs #1335 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…rivet, and report the trace-graph delta nobody had looked at The CHANGELOG's per-release section is hand-written prose about a typed set rivet can diff. v0.66 filed two of its OWN artifacts under [Unreleased] because the list was remembered rather than derived. scripts/release_notes_from_rivet.py extracts the previous tag's rivet sources with git archive — bounded by the paths rivet.yaml itself declares, so a moved source is a loud extract failure, not a silently narrower diff — runs rivet diff, and emits the artifact list AND the diagnostic delta. THE DELTA IS THE NEW PART. v0.69.0 -> this tree with the pinned 0.37.0: 8 added, 0 new ERRORS, 22 NEW WARNINGS in three classes shipped unseen — RQ-NN ids are not commit-trailer shaped (so rivet cannot trace them; the same convention WINDOWVAC had to teach status_evidence, found from the other side), five artifacts use a req-type the schema does not define, and none has an incoming verifies link. None fixed here; all now reported. THE VERSION CHECK IS LOAD-BEARING. This session's PATH rivet was 0.32.0 against a 0.37.0 CI pin (#1236/#1308), and 0.32 reports '0 broken cross-refs' on a tree where resolution never RAN — a clean bill of health from a check that did not happen. The generator refuses anything older than the CI pin, and test_pin_is_not_below_ci RE-DERIVES that pin from ci.yml rather than restating it, so the two cannot drift behind a green gate. Anti-vacuity: zero added artifacts REFUSES rather than emitting an empty section that would read as 'nothing shipped'. 8 unit tests, wired in ci.yml beside the loop-conformance and status-evidence potency tests. Refs #1337 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…the measurement and the question posted A deferral is not a delivery, so this deliberately does NOT wear the `RQ-NN-NAME (#issue):` delivery subject — RQ-70-WINDOWVAC teaches that shape to status_evidence's window in this same release, and R4 correctly refused the earlier subject as a delivery claim for an artifact whose status is `proposed`. The gate caught its author within the hour. Verified at source: refuse_custom_page_size and refuse_shared_memory live only in synth-cli/src/main.rs; synth-core's decoder decodes page_size_log2 and stops, its own comment saying it is "Consumed by refuse_custom_page_size in synth-cli". So a caller of the PUBLISHED synth-core still gets a custom page size silently ignored, and __synth_mem_size_N inherits the over-grant — the same shape #1317 closed in the same release on the opposite principle. NOT scoped here, deliberately. Pushing the refusal into synth-core forecloses a choice that is not mine: a 1-byte page is exactly the granularity gale's per-memory MPU region table (#1145) might WANT, and removing a mode before knowing whether an executed criterion needs it is what v0.68 nearly did to --safety-bounds mpu. Asked on #1145 today, three questions, none answered yet. The measurement is recorded so v0.71 need not re-derive it, and "CLI-only is correct" is an acceptable answer that keeps this deferred WITH a reason rather than dropped. Stays `proposed`, so #1315 is not closed on the tag. Refs #1315 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…erified at the cut, tags checked BEFORE not after
The conformance gate reported `DERIVED-FAIL: the only filed steps-1-2 decision is
RQ-63-ARCHMODEL, scoped to release v0.63 - NOT v0.70` while there was still time
to file. v0.69 learned this the other way round: its ARCHMODEL had LOST the
`feature-loop` tag, so the artifact filed to keep a deferral VISIBLE was itself
invisible to the gate, and the release read DOES-NOT-CONFORM until the cold
review caught it. Checked against the predecessor's tags this time, before the
cut.
Re-verified rather than remembered, observed 2026-09-18T21:13:53Z:
spar#445 state=OPEN updatedAt=2026-09-03T10:57:37Z comments=0
Unmoved since the v0.68 and v0.69 cuts. The count is the carried-from chain
walked back to the v0.63 head, not an asserted ordinal — release artifacts only
exist from v0.63, so "the Nth N/A" is not derivable and is not claimed.
A deferral filing is not a delivery, so this wears a `plan(...)` subject.
Refs #1136
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…, and guard the empty-areas case
NUMBERS, MEASURED AFTER THE REBASE ONTO THE MERGED ALIAS WORK, never added to
the previous figures (the v0.69 lesson: 856 vs the 854 measured before the
cascade settled, which `cargo fmt` then moved twice more):
selector_lines_code 19478 -> 19529 (+51, waived with the reason in claims.yaml)
selector_lines_total 30071 -> 30122 (tracked, no waiver needed)
home-alias audit 7203 -> 7207 functions, 8214 -> 8220 homes
(one new scripts/repro fixture x the two hard-float legs;
hits: 0 and 217081 attributed unchanged)
THE EMPTY-AREAS GUARD. `policed` answered `Some(areas) => areas.iter().any(..)`,
so `Some(&[])` would have disabled the #331 check entirely. Nothing produces it
today — select_with_stack collapses an empty list to None — but the degradation
belongs in the predicate rather than in one caller's care, and the load-bearing
question for any "unknown" value is whether it means POLICE EVERYTHING or POLICE
NOTHING. Pinned by ra003_areas_empty_is_unknown_not_permission_1321; ra003 is now
39 tests and RQ-70-ALIAS's artifact says 39 rather than 38.
RQ-70-DONEWHEN records its own measured outcome: manual-share went 100% for four
releases running to 22% here (7 of 9 done-when mechanically evaluable). Nothing
banned `manual:` — R9 refused RQ-70-ALIAS's first signature and the standard
spread. The two that remain manual are the two DEFERRALS, where no mechanical
predicate exists because nothing landed.
Refs #1321
Refs #1331
Refs #1335
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
… gap gets its own issue (#1339) A false statement in my own artifact, caught before the cold review and recorded rather than quietly edited. The first draft said "#1315 stays OPEN; this artifact is `proposed`, so the release does not close it." #1315 was CLOSED on the v0.69 tag at 2026-09-18T17:53:18Z by RQ-69-PAGESIZE, which shipped the CLI refusal. The REASONING was right — do not close an issue whose artifact claims no completion — and the FACT was wrong: I applied that rule to an issue another artifact had already closed for its own valid reason. So this is a follow-up to a CLOSED issue, not a deferral of an open one. It gets its own tracking issue (#1339) rather than a reopen, because the library-side gap is a different defect from the one v0.69 fixed, with a different fix site and a different consumer — and reopening a correctly-closed issue would misreport what v0.69 delivered. `fields.issue` now points at #1339. Refs #1339 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…them drifting
The first draft said "193 modules on main" — already wrong by one, because the
ALIAS fixture had merged in between. A stale number inside the artifact that
exists to say numbers go stale is the R11 blind spot in its purest form.
Counted with the SAME non-recursive glob the scripts use (a recursive count
differs again — scripts/repro has subdirectories):
at v0.69.0 A=193 B=209 x5 = 965 / 1045
at main today A=194 B=210 x5 = 970 / 1050
with this PR A=195 B=211 x5 = 975 / 1055
So the plan's "1035 = 207 x 5" was right for the tree it was measured on and
wrong two releases later, and "965" moves every time a fixture lands — including
the two this release adds. Showing the drift across three refs makes the point
the artifact is about far better than any single figure could.
Refs #1318
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
…ed output format and a YAML shape rivet rejects CI found both; neither was visible to any local gate, and both are instances of the class this release is about. 1. THE NON-VACUITY FLOOR READ A FORMAT I HAD MOVED. ci.yml asserted `^artifact citations: N cited filters over M test names`. RQ-70-CITEGAP rewrote that summary line to carry the target/filter split and the file count, so the grep matched nothing and the STEP failed while the GATE passed. Changing a gate's output means changing whatever reads it. The replacement floor is STRONGER than the one it replaces: it pins the artifact-FILE count above 100. The pre-fix non-recursive glob saw 30 files and zero release artifacts, so a regression to it now fails here. NEGATIVE CONTROL, run: reverting the glob gives "30 artifact files" and the floor refuses it — while the gate itself still exits 0, which is exactly why the floor exists. 2. A YAML SHAPE PyYAML ACCEPTS AND RIVET REJECTS. RQ-70-ARCHMODEL's long one-line `verified-by` was dumped as a multi-line SINGLE-QUOTED flow scalar. PyYAML round-trips it, so claim_check, status_evidence and the citation gate all passed locally — and the required `Rivet Validation` job failed with `RQ-70-ARCHMODEL.yaml:65: expected ':' after mapping key`. Reproduced in isolation (it needs the `common` schema loaded) and bisected to the field. Fixed by emitting FOLDED block scalars (`>-`), which carry the same logical single-line value and are what the hand-written artifacts use. Swept the CLASS rather than the instance: RQ-70-PAGELIB had two more multi-line single-quoted values that happened not to trip the parser, and a scan now reports zero remaining across artifacts/release-v0.70/. Refs #1333 Refs #1136 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
…s guard shifted liveness.rs again `reanchor` moved 5 sites, 0 not found — pure relocation (17 insertions / 17 deletions, only the location keys and reanchored_at change). test_mutation_survey and claim_check green on the re-anchored ledger; the `mutants_untested` ratchet does not move, because relocating a site changes where it is, not whether it is tested. This is the SECOND re-anchor of this release: the first followed RQ-70-ALIAS's ~50 added lines, this one follows the 12-line empty-areas guard. Both were correct and mechanical, and both cost a commit plus a full CI round — the ledger pins sites by file:line:col, so any edit ABOVE a pinned site in an anchored file invalidates it. Worth noting rather than fixing here: the same file already anchors its one borderline record by its `before` TEXT precisely because "line numbers move" (see BORDERLINE_LOUD_BEFORE and the v0.66 incident its comment records). The ledger itself could use that same stable anchor. Filed as a v0.71 candidate, not smuggled into a release fix. Refs #1189 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L
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
1 participant
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.
Six v0.70 artifacts in one PR, deliberately batched: they touch disjoint files (one Rust lane, four Python-gate lanes, one deferral), each is a separate commit so
status_evidence's R4/R10 attribution resolves per artifact, and every merge stales every other PR'sbaseRefOid— so batching costs one CI round instead of six.RQ-70-NPA (#1331) — the second
must, cpetigStatic-data
f32.load/f32.storenow relocate under--native-pointer-abi.The machinery existed —
emit_wasm_data_addrhas served the i32 sub-word (#744) and i64 (#746) arms since #739. The float arms declined loudly, and correctly: the raw[R11 + addr + #offset]path bakes the linmem offset as an un-relocatedMOVW/MOVT, invisible to the--shadow-stack-sizedown-shift and the post-link in-range oracle because both walk relocations. So this ships the relocation, not the bake. Both arms already bit-cast through a core register, so only the address changed and the bytes are identical by construction.Not claimed closed.
position#tickandekf#estimatenow reach a different ceiling the f32 decline was hiding:LdrSym literal pool out of range (#345): imm12=11744 / 4344 > 4095.imm12is the PC-to-pool distance, so 11744 is an ~11 KB function; each relocated access adds at most one 4-byte pool word, so even 50 added entries leave ~3× over the limit. Pre-existing, unmasked — not introduced.RQ-70-CITEGAP (#1333) — a gate that had never read a release artifact
Three defects, and the third was only visible because the first two were fixed:
run:keys; release artifacts cite underdone-when:/verified-by:.--test <target>names an integration-test file,--lib/--are substrings over test paths. Measured: 4 of v0.69's 5 citations are targets, so widening the glob alone would have produced four false failures.Red-first: planting
--test this_test_does_not_exist_9999intoimplementedRQ-69-ELFNAMES left the gate at rc=0 printing "every cited test filter resolves". After the fix the same plant reds, naming artifact, status and missing target.The widening then exposed a real scan bug — module paths were built only from
moddeclarations, never from the file, so a valid--lib dwarf_line::read as dangling. Fixed; test names 5233 → 7224.Keys are limited to
run/done-when/verified-by, notdescription: RQ-69-PROSEGATE refuted prose-scanning one release ago — it cannot tell a CLAIM from a MENTION.RQ-70-WINDOWVAC (#1334) — R10 was blind to this project's own convention
Re-derived per window with the script's own regex:
The RQ column rising to 8 exactly as delivery-shaped falls to 0 is the mechanism. The recognizer already existed (
ARTIFACT_ID, what R4 uses) and was simply never applied in the window loop. CI could not catch it: the grep is[0-9]+ delivery-shaped, and[0-9]+matches 0 — #1243's shape.Proven both ways: replaying v0.68.0..v0.69.0 now reports 8 delivery-shaped / 8 attributed (was 0); rewriting those subjects
RQ-69-→DELIVERED-69-fires the new floor.RQ-70-DONEWHEN (#1335) — declarations printed as evaluation
manual:share: v0.66 8/8, v0.67 7/7, v0.68 9/9, v0.69 9/9 — four releases at 100%, all printing the sameN/N done-whenline as a release with everything machine-checked.Not a ban: all nine v0.69 predicates are compound acceptance statements no single signature carries. But most hide a mechanical conjunct written in prose and read by nothing — five real test citations, which is CITEGAP's finding from the other side. On v0.69.0 the gate now prints
9/9 DECLARED — 0 mechanically evaluable, 9 manual: (R3 cannot fire on these). Verdict unchanged: still CONFORMS.RQ-70-RIVETNOTES (#1337) — derive the notes, and report the trace-graph delta
scripts/release_notes_from_rivet.pyextracts the previous tag's rivet sources (bounded by the pathsrivet.yamldeclares, so a moved source is a loud extract failure) and emits the artifact list and the diagnostic delta: 8 added, 0 new errors, 22 new warnings in three classes shipped unseen.The version check is load-bearing: this session's PATH rivet was 0.32.0 against a 0.37.0 CI pin (#1236/#1308), and 0.32 reports "0 broken cross-refs" where 0.37 reports "cross-refs NOT CHECKED" — a clean bill of health from a check that never ran. The generator refuses anything older, and
test_pin_is_not_below_cire-derives that pin from ci.yml rather than restating it.RQ-70-FALCONCORPUS (#1318) + RQ-70-ARCHMODEL (#1136) + RQ-70-PAGELIB (#1315)
EXPECTED_DECLINES. And "the corpus" turns out to name three populations across 13 scripts —*.wat(7),*.wat+*.wasm(5),*.wasm(1) — which is how1035 = 207×5and965 = 193×5were both "the corpus" in one release. Not unified (the globs differ for reasons); the census is the deliverable.OPEN / 2026-09-03 / 0 commentsat 21:13:53Z. Filed before the cut because the gate saidDERIVED-FAIL … NOT v0.70while there was still time — v0.69 found this only in the cold review, after its ARCHMODEL lost thefeature-looptag.synth-coredecodespage_size_log2and stops. Not scoped, because gale's per-memory MPU work (Multi-memory phase 2: per-memory MPU isolation is unreachable — --safety-bounds mpu and multi-memory have disjoint paths #1145) is what would consume it and a 1-byte page may be a granularity they want. Asked on Multi-memory phase 2: per-memory MPU isolation is unreachable — --safety-bounds mpu and multi-memory have disjoint paths #1145; staysproposed, so custom-page-sizes: (memory 1 1 (pagesize 1)) compiles silently with __synth_mem_size = 0x10000 — region table over-grants, no diagnostic #1315 is not closed on the tag.Also here
An empty-areas guard for RQ-70-ALIAS:
Some(&[])would have disabled the #331 check entirely. Nothing produces it today, but the degradation belongs in the predicate, not in one caller's care.ra00334 → 39.Numbers, all re-measured on the rebased tree (never added)
selector_lines_code19478 → 19529 (+51, waived with the reason inclaims.yaml);selector_lines_total30071 → 30122.functions audited7203 → 7207,homes watched8214 → 8220 — one new fixture × the two hard-float legs.hits: 0unchanged.Gates
cargo fmt --check,clippy --workspace --all-targets -D warnings,cargo test --workspace, ARM corpus sweep PASS, claim_check 75/75, status_evidence / version pins / oracle wiring / citation gate all exit 0.Refs #1331
Refs #1333
Refs #1334
Refs #1335
Refs #1337
Refs #1318
Refs #1136
Refs #1315
🤖 Generated with Claude Code
https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L