feat: canonical view for translation-suffix bodies (translation-bot-compat)

[johnlanni]

Summary

• Adds a canonical (strip-only) view for issue/comment bodies carrying machine-translation suffixes from the GitHub translate action (lizheming/github-translate-action), so issue-spec parsing, validation, digests, and write-path guards tolerate translated bodies without rewriting them.
• Implements change translation-bot-compat per the issue-spec governance flow:
  • Proposal: Proposal: translation-bot-compat #452
  • Design: Design: translation-bot-compat #453 (converged with independent review: SOUND — READY FOR IMPLEMENTATION, 4 rounds)
  • Implement + TASKs: Implement: translation-bot-compat #454 (TASK-454001/454002/454003, all confirmed)
  • Durable SPEC: SPEC-452001 on Proposal Proposal: translation-bot-compat #452 (capability issue-api-compatibility)

Why

The translate action appends a translated copy after an exact divider line, duplicating issue-spec markers, ## sections, and canonical JSON sections. Live breakage: issue create design predecessor validation (found 2 markers), exactly-once ## Choice Model / ## Answer / ## Question parsing, duplicate relationship edges, and digest churn on every bot re-translation (observed on higress-group/HiRoute #1/#9).

What (per TASK)

• TASK-454001 internal/model/translation.go: divider allowlist (single exact literal, line-anchored), fence-state guard mirroring internal/preview (including stop-on-unclosed-fence), single-divider rule (strip only when exactly one qualifying divider exists; two → fail-closed), StripTranslationSuffix / CanonicalView. 17-case table test.
• TASK-454002 canonical view at every consumer: FindMarker/ParseTypedComment (composed view), LogicalBody + TypedSectionList (strip-only, byte-identical passthrough for untranslated preview bodies), validatePhasePredecessor, locator.parseMarker/exactPredecessor, canonicalProposalContribution; CanonicalTitle || normalization at exact-title comparisons.
• TASK-454003 RepresentationDigest re-based on the strip-only view (doc contract: "digest of the stripped raw body, no preview masking"; preview-source-only edits stay digest-visible); write-path exact-byte guards at engine.go and link.go become canonical comparisons with canonical-digest messages (suffix-only drift passes, canonical drift still conflicts); relationship semantic scans bounded at the divider (within-prefix duplicates still ErrAmbiguous); generated skill guidance note (never hand-strip, never author the divider literal outside fences).

Test plan

• go build ./...
• go test -p 1 ./... — 96 packages, 0 failures
• Design tests 1–10 each mapped to named tests (translation_test.go, translation_parsing_test.go, issue_translation_test.go, locator_translation_test.go, engine_translation_test.go, link_translation_test.go, mutation_translation_test.go, artifacts_translation_test.go, translation_workflow_test.go, title_test.go, model_test.go)
• Untranslated bodies remain byte-identical passthrough (no divider → no behavior change)

Design deviations (documented in the implementation report): stripped results canonicalize trailing newline runs so digest(original) == digest(original+suffix); D8 equality uses model.CanonicalBodyEqual (canonical views, join-byte tolerant); skill byte budget 12400→12800. Follow-up candidate noted: ObserveAcceptedReceiptAuthority still reads the raw semantic view (out of D1 scope).

Closes #452

[johnlanni]

Independent Review — qoder-k3 (Kimi-K3, exact head 280b607)

(GitHub forbids approving your own PR, so this is posted as a COMMENT; treat it as the review verdict.)

Verdict: APPROVE — D1–D11 faithfully implemented; all declared deviations verified as bounded; fail-closed behavior preserved everywhere the canonical view does not apply. No P0, no blocking P1.

Checklist verification highlights

• translation.go: allowlist single exact literal; fence opener/closer byte-semantics identical to preview.parseOpener/isCloser (≤3 leading spaces, ≥3 run, backtick-in-info rejection, stop-on-unclosed-fence); single-divider fail-closed rule; CRLF handled per preview splitLines.
• Integration completeness: every design-named consumer converted — marker.go:28, typed_comment.go:282, validation.go:39 (LogicalBody), links.go:139 (TypedSectionList), issue.go:333, locator.go:154,180, service.go:421, bindings/service.go:647-652, RepresentationDigest (typed_comment.go:26), engine.go:570,588,638,653, link.go:313,335,347, mutation.go:336. Remaining raw paths are same-suffix comparisons (transition.go) or fail-closed (see P2-2).
• Untranslated passthrough: byte-identical when no qualifying divider; pinned by tests incl. preview-body passthrough and the retained exact-bytes digest contract for no-suffix bodies.
• D8: all 7 sites use CanonicalBodyEqual; digest-equality implies view-equality, so no identical-digest-mismatch message is possible; real conflicts still reported with differing digests (pinned by TestReconcileStillConflictsOnCanonicalDriftAfterWrite); version-CAS residual documented at engine.go:561-563.
• D9: raw index arithmetic unchanged; suffix excluded; within-prefix duplicates still ErrAmbiguous; suffix survives mutation writes.
• Digest re-basing: all consumers route through the same canonical function — intra-run consistent; pre-change sealed digests of untranslated bodies unchanged; translated-body flows were already broken pre-change. Migration acceptable.

P2 findings (non-blocking)

1. Digest-stability boundary narrower than deviation-1 wording: digest(original) == digest(original+suffix) holds only when the original ends in exactly one LF (translation.go:28-33). Zero or ≥2 trailing newlines → mismatch at sealed-digest sites (engine.go:622,726-727,752-753, native_answers.go:196, link.go:335) — all fail closed, never corrupt. Recommend documenting the exactly-one-LF boundary and pinning a fail-closed test for a no-trailing-newline original.
2. Scope follow-ups (fail-closed, same class as deviation 4): parseProjectionMarkers (projection.go:555) and specialized marker parsers (rationale.go:28, finding.go:52,140, code_change_rationale.go:229) still scan raw bodies. Recommend one tracking issue bundling them with ObserveAcceptedReceiptAuthority.
3. Perf nit: add a strings.Contains(body, divider) fast path before split/join on the dominant no-divider path.
4. Unpinned adversarial cases (verified correct by reading): blockquoted divider, mixed fence chars, divider-as-last-line, 4-space-indented divider (currently qualifies per "optional surrounding whitespace" — intentional).

Reviewed HEAD: 280b607b6570536c5346742a973bdad36dacd801. Worktree clean throughout; static analysis only (in-session test execution was unavailable — suite green per author and CI).

[johnlanni]

Re-review Round 2 — qoder-k3 (exact head 53fd5d7)

Verdict: APPROVE — all four P2 responses implemented as described, correctly bounded, pinned by non-tautological tests.

• P2-1 boundary doc accurate (translation.go:23-32); both directions pinned (TestRepresentationDigestStabilityBoundaryOnTrailingNewlines: one-LF equal / zero-LF and 2+-LF mismatch; link.go:335 real fail-closed path with 0 writes). Documenting rather than manufacturing stability is the right call — always-normalizing would break sealed-digest continuity for untranslated bodies.
• P2-2 conversion choices match each site's masking precedent (composed for projection.go:555 + typed_comment.go:174 receipt authority; strip-only for the four raw marker parsers); no byte-offset mapping back to raw bodies, so no index desync; genuine duplicates still fail exactly-once. Closes round-1 deviation 4.
• P2-3 fast path is pure perf: strings.Contains gate returns identical string/flag; AllocsPerRun=0 pinned.
• P2-4 four adversarial pins assert correct behavior (blockquote rejected, mixed fence chars → unclosed → fail-closed, divider-as-last-line recovers original exactly, 4-space indent qualifies per spec'd 'optional surrounding whitespace').
• Regression hunt clean; write-side preflightProjectionBody correctly left non-canonicalized (guards locally authored files). DCO/author consistent (澄潭 zty98751@alibaba-inc.com; the 'alibabainc' typo was report prose only, not the commit).

Caveat: static review per session constraints; suite green per implementer report and CI (96 packages ok).

Reviewed HEAD: 53fd5d7