Regression Verification Audit

Confirm that previously-fixed bugs are still fixed. This audit dynamically
discovers closed bug issues from GitHub, locates each fix and its guard test,
and reports any fix that has gone missing as a Regression of #NNN.

Read _audit-common.md (dedup, methodology, per-finding format) and _audit-severity.md for shared protocol. This file only adds the regression-specific flow.

Parameters (from $ARGUMENTS)

• --issues <N,N,N>: verify exactly these issues (skips discovery).
• --limit <N>: max issues to verify (default 40).
• --label <label>: issue label filter for the fallback pass (default bug; use bug,documentation,doc-rot to include doc-rot fixes, or game:<title> to scope to one title).
• --recent: skip churn weighting; take the N most-recently-closed issues instead.

Step 1 — Discover fixes worth re-checking (churn-weighted)

A fix can only regress if the code around it changed. With thousands of closed issues, "the last 50 closed" re-checks fixes nobody has touched and never reaches old fixes in hot files. Select by churn instead:

mkdir -p /tmp/audit
D=$(ls docs/audits/AUDIT_REGRESSION_*.md | sort | tail -1 | grep -oE '[0-9]{4}-[0-9]{2}-[0-9]{2}')   # last sweep
base=$(git rev-list -1 --before="$D 23:59" HEAD)
git diff --name-only "$base"..HEAD -- '*.rs' '*.glsl' '*.comp' '*.frag' '*.vert' > /tmp/audit/churn.txt
# earlier fix commits on the files that changed since, ranked by overlap:
xargs -a /tmp/audit/churn.txt -I{} git log "$base" --format=%s -- {} \
  | grep -oiE '(fix|fixes|fixed|close[sd]?|resolve[sd]?) #[0-9]+(( *(,|\+|&|/|and) *| )#[0-9]+)*' \
  | grep -oE '[0-9]+' | sort | uniq -c | sort -rn | head -"${LIMIT:-40}" > /tmp/audit/candidates.txt

The (…#N)* tail is load-bearing: multi-issue subjects (Fix #5075 + #5076: …, Fix #5215, #5216: …, Fix #1297 / #1298) are common, and a single-#N pattern drops every issue after the first — 11 of 279 fixed issues in the 2026-09-29 → 10-05 window. A repeated keyword (Fix #5393, Fix #5395) matches as two hits; a range (Fix #5371-5381 batch) yields only its first number, and nothing is lost — the two range subjects in history are batch follow-ups, and each of #5371–#5381 has its own Fix #N commit. Issue numbers cited later in a subject (… catches up with #5154) are deliberately not matched.

Take the top candidates, then gh issue view <N> --repo matiaszanolli/ByroRedux --json number,title,body,closedAt,labels for each. If the churn list is empty or short, top up with gh issue list --repo matiaszanolli/ByroRedux --state closed --label bug --limit 50 --json number,title,body,closedAt,labels. Never trust a hand-typed closed-issue count; ask the API.

For each issue pull out: number + title; file references (backticked paths in the body); the fix description / acceptance criteria; related #NNNN — phased fixes (e.g. #1210 → #1255 → #1257) regress as a set, so verify the whole chain. Note #1651 was itself a wrong fix, disproven and reverted by #1823 — do not verify it as if it still holds.

Step 2 — Locate each fix and its guard

For each issue, work the fix → guard-test chain:

1. Find the fix commit. git log --oneline --grep="#<N>" (commits use
   Fix #<N>: … and fix/<N>-… branch merges). git show <commit> --stat
   shows which files moved.
2. Confirm the fix is present in the live tree. Read the referenced file(s)
   at the symbol named in the issue/commit (prefer grep -n "fn <name>" over a
   line number — the post-Session-34/35 module splits invalidate old line refs).
3. Find the guard test. Tests live as *_tests.rs siblings next to the
   module they cover (e.g. crates/nif/src/blocks/interpolator_tests.rs), or as
   #[cfg(test)] mod tests inline. To locate one:
   • grep -arn "<N>" crates/ byroredux/ --include='*.rs' — many tests cite the
     issue number in a name or comment (fn fix_1516_…, // #1516).
     The -a is load-bearing (#3210): a single raw NUL byte anywhere in a
     .rs file — trivially introduced by writing b"Foo<NUL>" where the source
     meant b"Foo\0", and accepted silently by rustc — makes plain grep
     classify the whole file as binary and skip it without a word. That once hid
     40 guards citing 31 issues in one 1,944-line file, and the sweep that found
     it came one command from filing a FAIL against a fix that was present.
     rg and git grep are immune; plain grep is not.
     scripts/check-text-source-integrity.sh (run in CI, and by
     _audit-validate.sh) now rejects such files, but keep the -a: a
     PARTIAL/FAIL published on a blind grep costs far more than a flag.
   • Failing that, grep the fixed symbol or a keyword from the title across
     *_tests.rs siblings of the fix file.
4. Run the guard to prove it still passes. Crate packages are named
   byroredux-<crate> (e.g. cargo test -p byroredux-nif <test_name>,
   cargo test -p byroredux-renderer, cargo test -p byroredux-core --features inspect
   — without inspect the animation-serialization guards silently drop, #3895).
   The byroredux bin crate (and any workspace-resolving call) needs rustc ≥ 1.96 —
   rustup cargo per docs/contributing.md § Toolchain note (#4466). Never run
   --ignored on byroredux-plugin (real-master tests OOM past 20 GB).

Step 3 — Assign a status

• PASS — fix code confirmed present and a guard test exists (ideally run green).
• PARTIAL — fix code present but no guard test. Flag as a hardening gap.
• FAIL — fix code missing or its guard now fails. This is the regression.
  Report it with Status: Regression of #<N> per the _audit-common finding format.
• UNVERIFIABLE — the issue body names no file/symbol and no fix commit is
  findable. Note it and move on; don't guess.

Step 4 — Unconditional fragile-area guards

These contracts landed as refactors, not closed bugs, so issue discovery never surfaces them. Run every time; a failure is reported as a regression (cite the issue if one exists, else the contract). Deep checklists live in the owning audit — this step only proves the guards are live.

Area	Contract	Verify
GPU struct sizes	every #[repr(C)] shader-contract struct keeps its pinned size (the pins, not this row, hold the numbers)	cargo test -p byroredux-renderer gpu_ (crates/renderer/src/vulkan/scene_buffer/gpu_instance_layout_tests.rs, crates/renderer/src/vulkan/material_tests.rs)
NIFAL single boundary	translate_material in byroredux/src/material_translate.rs is the only ImportedMesh → Material site; metalness/roughness stay plain resolved f32; Material::resolve_pbr fills only unresolved slots	cargo test -p byroredux-core resolve_pbr; /audit-nifal
Typed particle emitters	NiPSysEmitter* parse typed (crates/nif/src/blocks/particle.rs) → extract_emitter_params (crates/nif/src/import/walk/emitter.rs) → apply_emitter_params (byroredux/src/systems/particle.rs); an opaque NiPSysBlock shows as zero-sized emitters	cargo test -p byroredux apply_emitter_params
Collision coverage	every Bhk*Shape parser maps to a CollisionShape in crates/nif/src/import/collision/shape.rs (incl. BhkMultiSphereShape, BhkConvexListShape)	cargo test -p byroredux-nif collision
ReSTIR reservoir	the per-thread resRadiance[] array stays retired (#1369) — WRS is register-local, radiance recomputed via shadowableLightRadiance in crates/renderer/shaders/include/lighting.glsl; verify it stays gone, not "intact"	grep resRadiance in crates/renderer/shaders/ → only the retirement comments
Scheduler access	every parallel system declares what it acquires	cargo test -p byroredux system_access_declaration_tests
Save shape	serialized shape changes force a baseline refresh or FORMAT_MAJOR bump	cargo test -p byroredux serde_default_guard_tests

Output

Write to: docs/audits/AUDIT_REGRESSION_<TODAY>.md (YYYY-MM-DD).

Per-issue entry

## #<ISSUE>: <Title>
- **Status**: PASS | PARTIAL | FAIL | UNVERIFIABLE
- **Closed**: <date>
- **Fix commit**: <hash> (or "not found")
- **Fix site**: `<path>` (`<symbol>`)
- **Fix present**: Yes / No / Unknown
- **Guard test**: `<test name>` in `<path>` — passes / fails / none
- **Notes**: <concerns>

Summary table

| Issue | Title | Status | Fix Present | Guard |
|-------|-------|--------|-------------|-------|

For any FAIL, surface it as a Regression of #NNN finding (base format in
_audit-common.md) and suggest:
/audit-publish docs/audits/AUDIT_REGRESSION_<TODAY>.md