fix(harbour): strengthen repair reliability from real Anna reruns

[zedarvates]

Summary

• preserve the privacy-safe evidence from the rejected cross-model experiment and the complete real Anna corpus runs;
• keep the fail-closed A06 evidence showing that missing creative structure cannot be reconstructed deterministically;
• strengthen the single allowed model-repair pass with a schema-complete required-field checklist and hard structural invariants;
• retain regression coverage for exact input preservation, private-response exclusion, canonical validation, forced colours, browser flows, and Marketplace-logo paths;
• document the direct Host LLM structured-output boundary without adding an Executa or changing the Host-API-only MVP architecture.

Current measured evidence

• Anna default, MiniMax M3 through OpenRouter on 27 August: 6/20; median 49.13 s; p95 57.87 s; 2 repaired passes;
• Gemma advisory preference before the schema-complete repair on 27 August: 15/20; median 21.91 s; p95 41.92 s; 6 repaired passes;
• Gemma after the schema-complete repair on 31 August: 15/20; median 23.90 s; p95 44.00 s; 7 repaired passes;
• post-repair failures: HBR-A01 unknown after one transport failure, plus HBR-A06/A08/A15/A20 json_invalid;
• a bounded offline punctuation-only experiment could recover only two malformed cases, projecting at most 17/20, and was removed.

The authoritative real-model gate remains below the required 18/20. No score improvement is claimed from the repair change.

Reliability change

The repair request explicitly enumerates every required project, production-bible, character, location, scene, shot, continuity-report, and warning field. It reiterates the three-scene/one-shot budget, reference integrity, ordering, array/string types, warning severities, continuity-score bounds, original-input preservation, and 12,000-character ceiling.

The repair still:

• discards the previous malformed/private model response;
• rebuilds only from the original user input plus privacy-local validation errors;
• gets exactly one attempt;
• remains subject to the unchanged canonical validator;
• fails closed if required creative structure is absent or JSON remains invalid.

Structured-output boundary

Anna currently documents json_object and json_schema for Executa sampling, not for direct iframe anna.llm.complete. Pinned CLI 0.1.30 and read-only inspection of CLI 0.1.49 expose no supported direct Host API response-format field. Moving Harbour into an Executa solely for structured output would violate the approved Host-API-only MVP boundary.

review/ANNA_STRUCTURED_OUTPUT_REQUEST.md contains the bounded support question and one-prompt validation gate. Nothing has been sent, no package was upgraded, and no speculative model call was made for that investigation.

Validation

• StoryCore Harbour CI: green on head e076230;
• SonarQube Cloud Quality Gate: green;
• PR merge state: clean;
• private JSONL and harness logs remain ignored and are not attached to this PR.

Release boundary

This remains a draft PR. It does not authorize an immutable Anna version, App Review submission, merge, installation claim, or publication. StoryCore Harbour remains a mutable Anna draft with zero immutable versions. The immutable release gate remains >=18/20 with median successful completion <=180 seconds on the fixed corpus.

[zedarvates]

Reliability patch ready for real rerun

The repair path has now been strengthened from the measured 2026-08-27 failures without changing the immutable corpus or relaxing the canonical validator.

Changed

• repair requests now enumerate every required project/bible/entity/scene/shot/continuity field;
• hard rules explicitly cover array/string types, reference integrity, exact scene/shot counts, ordering, warning severities, score bounds, and preservation of the original input;
• the previous malformed/private model response is still discarded;
• only one repair attempt remains allowed;
• missing creative structure still fails closed rather than being invented deterministically;
• regression tests lock this behavior.

Automated result

GitHub Actions on head 234987a796485083bfd9dd14d881656a9b7a817b passed the full Harbour job: syntax, Node tests, sample export, mock fixture, immutable corpus, generated-bundle synchronization, strict Anna validation, browser flow, and Marketplace screenshot generation.

Next evidence to collect

Do one complete Gemma-advisory corpus rerun first, because the latest Gemma baseline is the closest to the gate (15/20, median 21.91 s) and its five failures are structural/required-field failures targeted by this patch.

Do not claim an improvement from this patch until that complete result exists. Required release gate remains >=18/20 and median <=180 s. If Gemma reaches the gate, repeat once to check stochastic stability before any readiness claim. If it remains below the gate, record only privacy-safe failing prompt IDs/categories and continue from measured evidence.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud