Align served Gold bodies with signed canonical line endings

[patschmittdev]

Summary

Align served Gold body line endings with the existing signed canonical representation. This is a representation-consistency fix, not a cryptographic collision or admission bypass. Based exactly on audited commit a62686c3ab6d58c67f7b1c58529436099372b95d.

makeGoldChunk in src/retrieval/chunks.ts reuses the existing CRLF-only helper for the parsed source representation, then authorization's existing normalizeText before constructing the Gold body and body-derived chunk ID. Review/evidence Gold, BM25, citation hashes, and integrity fingerprints consistently consume that canonical LF body.

Normalizing body and ID alone would hide a post-build lone-CR mutation from freshness checks. A host-derived, optional source_body_sha256 cache field preserves the parsed pre-canonical body hash only when it differs from canonical LF. Gold stores it on the chunk, and review/evidence copy it into authorization provenance. Constructor and schema field order match. Removed, tampered, and unexpected markers fail verification, including when cache fingerprints are recomputed.

Compatibility: This is a bounded optional index-v2 cache schema extension, not a receipt change. Ordinary LF/CRLF serialized chunks, IDs, and fingerprints remain unchanged. If indexed Gold contains lone CR, rebuild all three indexes with ziggurat build after updating. Those legacy snapshots fail live verification; rebuilding changes their Gold chunk IDs, citation body hashes, and fingerprints. Obtain new citations afterward. Rolling back to an older binary also requires rebuilding affected indexes because older strict schemas reject the new field. No re-signing or Silver restaging is needed.

Changed files:

• src/retrieval/chunks.ts
• src/contracts/gold-index.ts
• src/retrieval/profile-index.ts
• test/gold-body-normalization.test.ts
• docs/retrieval-evaluation.md
• site/src/content/docs/concepts/tiers.md

Trust-boundary impact

Only host-constructed Gold cache representation and related cache provenance change. The source-body marker is neither an admission nor signing input. Receipt v1 canonical bytes, digest algorithm, signing payload, published vectors, and verification are unchanged. Generic corpus parsing and Bronze storage/canonicalization are untouched: Bronze still normalizes CRLF only and preserves lone CR in body hashes and line-range citations. Source files, receipts, and trust configuration are not rewritten. Machine JSON remains lossless with no double escaping.

All-profile LF -> lone CR -> mixed -> LF mutation tests retain stale-search/citation refusal before rebuild, recover canonical retrieval after rebuild, and invalidate citations issued for the prior corpus state. Three physically separate indexes and shipped Gold-only/read-only MCP remain unchanged. No signer, admission bypass, trust writes, or model authority are added.

Validation

• Before implementation, the initial dedicated suite passed 6/9 tests and failed the lone-CR, mixed-ending, and post-rebuild canonical-body assertions. Existing pre-rebuild mutation refusal passed.
• Final dedicated suite: 13/13 passed with real temporary vault files and existing disposable signing helpers, without filesystem mocks, production keys/data, or model calls.
• Authorization-vector, authorization, retrieval, retrieval-quality, index-boundary, live-revocation, and dedicated suites: 94/94 passed.
• npm run check: passed, 708 core tests passed and 2 existing permission-dependent tests skipped because this filesystem/user does not enforce those permissions; all 11 script tests, links, and style checks passed.
• npm run site:check: passed, zero Astro diagnostics, 7 plugin tests passed, all 24 built pages validated.
• node dist\src\cli\main.js check --root . --audit-clean-room: PASS, no findings and no exclusion changes.
• git diff --check: passed.

This PR requires independent human review and the normal required Actions checks. No merge or protection bypass is requested.

Checklist

• Behavior changes include tests that use real temporary files.
• Model pathways still cannot write Bronze, knowledge, trust, receipts, reviewed metadata, or indexes.
• No signer, apply, approve, promote command, or --promote flag was added.
• Retrieved content remains reference-only with instruction_authority: none.
• Documentation and compatibility notes match the implementation.
• npm run check passes.
• node dist/src/cli/main.js check --root . --audit-clean-room passes.
• No private key, credential, personal path, or private vault content is included.