Bug-hunt round 41: analyze schema-rule doc gap, -compute_type openness marker

[REPPL]

Round 41 of the autonomous bug-hunt loop (state: issue #24).

Confirmed substantive (1)

.abcd/development/brief/04-surfaces/06-analyze.md's ingest section claimed to run "every schema rule" and silently omitted two.

The page lists six rules analyze -ingest enforces, introduced as "every schema rule" and closed with "Validation is transactional and exhaustive." It omits two rules internal/analyze/validate.go actually enforces unconditionally when present:

• the mode field's A|B enum (internal/analyze/validate.go:164-165)
• the 64-id cap on evidence (internal/analyze/validate.go:17, :174-175)

This is the identical defect round 39 already fixed on the sibling user-facing page, docs/reference/cli.md:156 ("the type, severity, and mode enums, non-empty evidence of at most 64 ids…") — round 39 fixed only that page and never swept this one. The brief page also contradicts its own cited source, .abcd/development/brief/05-internals/02-schemas.md:97,99, which documents both rules. Fixed to name both, matching cli.md's wording.

Confirmed nitpick (1)

docs/reference/cli.md's transcribe synopsis and internal/cli/cli.go's usage banner rendered -compute_type as a closed set, unlike its actual pass-through behaviour.

Both render [-compute_type auto|int8|float16] in the identical pipe-alternated form as three genuinely closed, validated siblings on the same line — -engine, -device, -vad, each refused at exit 2 by CheckEngine/CheckDevice/CheckVAD (internal/cli/cli.go:254,261,264). -compute_type has no validator anywhere in the tree and is a deliberate pass-through to whisperx (internal/cli/cli.go:228-230's comment states this explicitly; internal/transcribe/whisperx.go:96-103's resolveCompute only substitutes for ""/auto). Reproduced live on the built binary: -device bogus, -vad bogus, and -engine bogus all exit 2 with a usage error; -compute_type bogus and -compute_type float32 do not — they reach engine detection. float32 is in fact a real, working CTranslate2 compute type the CLI's own source names (internal/transcribe/transcribe.go:42), so the synopsis both over-promises validation and under-advertises a working value.

Both sites now carry the openness marker (|…), matching the pipe form already used correctly at docs/how-to/transcribe-a-recording.md:80 and the flag table at cli.md:59.

Considered and refuted

• internal/cli/cli.go:183's -compute_type help string using ASCII "..." where cli.md's prose table uses "…". Split verdict on adversarial verification: one refuter found no stated repo convention ties code-side text to doc-side typography (the "US English in code" rule is about spelling, not glyphs), and that a single-line fix would desynchronise the help string from the identical ASCII list in its own explanatory comment at cli.go:229 — a coherent fix would need both lines, making the finding as originally scoped incomplete. Discarded per the loop's tie-breaking rule.

Verification

Every finding faced two independent adversarial refuters before being included above; only findings both refuters failed to kill were fixed. Dimension (a) code bugs and dimension (c) infrastructure surfaced zero findings this round — 40 prior rounds have hardened both areas heavily.

Gates

go build, gofmt -l ., go vet ./..., go test ./..., go test -race ./..., the pipeline smoke (merge + report against examples/sample-session), and sh -n install.sh && bash -n install.sh all pass on the branch head. Both fixes are documentation-only edits with no behavioural code change, so no new tests were required.

.abcd/work/DECISIONS.md gains the round's entry in this PR.

---

Assisted-by: Claude:claude-sonnet-5