feat: persisted, regression-gated interop matrix (§17 Requirement 18)

[SoundMatt]

Summary

Closes #145 (NEW-SPEC-7). Extends §20.1's behavioural-conformance gate
and the relay interop command so cross-language equivalence is a
persisted, diffable, regression-gated artifact instead of a one-shot
CI check whose result vanishes after the run.

Background

§20.1 item 3's existing gate checks EQUIVALENT-vs-reference every run,
but the result is never committed anywhere — a cross-language
equivalence regression has no artifact trail, only whatever the CI log
happened to say at the time.

What's added

• relay interop's JSON output is now a versioned artifact:
  kind: "relay-interop-matrix", matrix_version: "relay-interop-matrix/1", and an explicit participants[] list —
  so the set of compared implementations is itself part of the
  diffable artifact, not just the per-cell verdicts.
• New --baseline FILE flag: compares a fresh run against a
  previously-committed matrix and reports regressions[] — one entry
  per (vector, participant) cell that was EQUIVALENT in the baseline
  but isn't now. A cell that was never EQUIVALENT to begin with
  (e.g. an implementation still catching up to a new golden vector) is
  not reported as a regression — matches the issue's precise
  wording, distinct from the existing any-mismatch-fails gate.
• New §17 Requirement 18: CI MUST commit the matrix (covering every
  published sibling it can invoke) and MUST regenerate + --baseline
  it on every change, failing the build on any regression.
• New spec/schemas/relay-interop-matrix.json, new REQ-RELAY-101.

Deliberately not pairwise (N²)

The issue's literal wording ("compare against every other sibling
implementation, not only the reference") would mean diffing every
binary's convert output against every other binary's directly. Not
built: byte-equality is transitive, so "A equivalent to the shared
reference" and "B equivalent to the shared reference" already jointly
establish "A equivalent to B" — the existing O(N) hub-spoke comparison
carries the same information an O(N²) pairwise matrix would, without
the redundant convert invocations. Documented explicitly in §17 and
§20.1 rather than silently narrowed.

Bug found along the way

§20.4 cited §20.1.2 for a specific CI-gate item, but §20.1 is a flat
numbered list, not subdivided into .1/.2/.3 subsections — the
same class of citation drift the issue itself flags for the audit's
stale §20.1.3. Corrected to "§20.1 item 2".

Version

SpecVersion 2.7 → 2.8 (MINOR — new §17 requirement).

Verification

• go build/vet/test clean, gofmt -l . clean, golangci-lint run 0
  issues
• gofusa check 0 errors/PASS, gofusa trace -req-coverage 100
  100%/100% (101/101 requirements traced)
• Manual end-to-end: built binary, ran relay interop --format json
  against itself, fed the output back as --baseline — clean
  round-trip, no false regressions; relay conform --manifest reports
  18 total requirements with Requirement 18 correctly
  NOT_OBSERVABLE/implementation's own CI
• 9 new/changed interop tests, including a real mutation test on the
  regression detector's core comparison logic — caught immediately by
  a pure unit test on interopRegressions; the first
  integration-level attempt was fooled twice before correctly
  discriminating the mutation (the reject-path vectors' by-design
  leniency toward a broken convert, then a substring collision
  between the built test binary's name and the ever-equivalent relay (reference) row) — both false-pass paths fixed and reconfirmed.