plan(v0.62): scope the release — "Reach is part of correctness" (7 artifacts)

[avrabe]

v0.62 — "Reach is part of correctness"

v0.61 was 15 of 15 artifacts fixing checkers. That was the right scope for
one release and the wrong scope for two. CLAUDE.md's North Star extension says
it directly — "A proof about input we refuse is worth nothing" — and the
measured acceptance on 805 real modules is ARM 66%, RV32 14%, AArch64 1.6%.
So v0.62 turns outward.

The opener is not a defect we found

#1131. A real integrator is running a 5-stage --relocatable cascade on an
emulated RT1176. The whole documented remedy chain lowers cleanly, exit 0. They
stopped at the one part of the contract we never wrote down — R9 is specified,
the linear-memory base is not — and they asked rather than reverse-engineering
it from disassembly.

A user blocked on our docs outranks anything on this tracker that we found
ourselves.

Scope

priority	artifact	issue
must	RQ-62-EMBEDDER	#1131	the --relocatable layout contract, derived from the emitter — a hand-written contract that drifts is the class #1080 cost 98 files
must	RQ-62-TABLEDANGLE	#1102	the only v0.61 residual that is a correctness hole
must	RQ-62-REACH	#242	VCR-REACH-002, scoped for the first time; increment 1 is re-measurement, not implementation
should	RQ-62-MVLOWER	#1093	param-block types decline everywhere but are still not lowered
should	RQ-62-CFOBLIG	#1057	Block/End correspondence — where all 180 of gale's remaining instances live
should	RQ-62-CLAIMCHECK	#1062	last required context with no system dependency still on the hosted quota
should	RQ-62-ROADMAPGATE	#1133	new, found while consolidating this scope

Two of these deserve calling out

RQ-62-TABLEDANGLE is a live correctness hole, not process debt. #1102's
guard matches direct-call index labels (func_{idx}/synth_func_{idx}),
which is complete for direct calls. A funcref table entry naming a declined
function has no direct call site, so no relocation carries its index label — the
driver-level gate sees nothing and the pre-fix unlinkable behaviour survives on
exactly that path. #1102's own history is that the same class was fixed on
aarch64 and left live on RV32 for a release because nobody checked the others.

#1133, found while writing this scope. status_evidence_check globs
artifacts/release-v* only, so the VCR roadmap's 39 artifacts are outside
every status/evidence rule — and two (VCR-WCET-001, VCR-WCET-002) have no
status key at all. Same class as v0.61's entire scope, one directory over:
the gate exists, is required, is green, and its scope doesn't cover the file
whose purpose is recording programme status. Deliberately not a glob
widening — roadmap items carry release: None by design, so the release-scoped
ruleset would produce false reds.

Residuals promoted, not left in prose

Every carried-forward item is a residual a v0.61 artifact named, moved from
prose inside a closed artifact to a tracked requirement. A residual recorded
only in a closed artifact's verified-by is invisible to release planning.

RQ-62-MVLOWER is tractable now only because #1097's oracle exists: it is the
acceptance test, and the guard may be relaxed only for a shape whose lowering
makes the four silent-wrong vectors correct.

Not an artifact, on purpose

#1132 (org-wide feedback request) is addressed to the maintainer. Answering
it is the deliverable; a rivet artifact would be the wrong response.

Gates

ARTIFACT_FLOOR 509 → 516, re-derived with rivet list, not computed.
status_evidence 0 failures · claim_check 56/56.

Refs #242, #1057, #1062, #1093, #1102, #1131, #1133

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!