[G3] Cross-repo parity acceptance and contract-drift CI: build burn against relayhistory main, run the corpus parity suite, publish a real-machine parity report

[willwashburn]

Part of #160 (group 3 — depends on all of group 2 here and burn #557/#558; parallel with #184).

Problem

Once burn depends on the published ai-hist crate, relayhistory can break burn without noticing: a parser change that alters message_id derivation, timestamps, or usage dedup silently changes burn's turn_content_fingerprint and either duplicates or collapses cost rows (burn ledger/fingerprint.rs:97-115). "Tests pass here" is a return value, not an effect; the effect lives in burn's ledger.

Scope

1. Contract-drift job in ci.yml: check out AgentWorkforce/burn (main), point its ai-hist dependency at this PR's workspace via [patch.crates-io], and run cargo test -p relayburn-sdk --features relayhistory-source --test relayhistory_parity (the suite from burn #557). Required check on PRs touching crates/ai-hist/src/** (after [G1] Merge ai-hist-core and ai-hist-engine into one published ai-hist crate with feature gates, a release pipeline and a semver policy #162 that is the whole library). A failure names the corpus case and the differing table/field.
2. Reverse job in burn (file the mirror task on burn #562's checklist): burn CI runs its parity suite against the published crate version it pins, so a relayhistory release cannot be adopted with drift.
3. Real-machine parity report: docs/parity/2026-XX-XX-burn-parity.md produced by running, on at least three developer machines with real Claude/Codex/OpenCode histories (and one with Cursor/Grok), both backends into separate ledgers and burn state parity --against (burn #555). Report: sessions compared, turns compared, exact-match %, every class of difference with root cause and the issue that fixes it, token totals per model and per harness for both, cost delta. Follow CLAUDE.md: confirm against effect (query both ledgers for specific message_ids), include a positive control (a deliberately altered turn shows up as a diff).
4. Corpus completeness: every difference class found on real machines that the corpus did not cover gets a new anonymized fixture in tests/fixtures/ (scrub content; keep structure) so it cannot regress.
5. Gate: [G3] Retire duplicated harness knowledge: docs, sibling-repo runbooks, burn collector transfer, and the "one owner" guardrail test #184 and burn #562 (cutover) may not start until this report shows 100% identity-vector match for Claude and Codex and every remaining difference is documented as intended.

Acceptance

• The drift job is a required status check and has caught at least one injected regression in a draft PR (e.g. change Codex token_json delta → parity fails).
• The parity report is merged with the numbers above and links to fixtures added.

Out of scope

Fixing what the report finds (goes back to the owning issue).