Supersede the legacy promotion to retire a stale expected failure

[DavidBakerEffendi]

Closes #164.

cpp-parity-function-like-macro-expanded-call carries an expectedFailure
annotation that Bifrost no longer earns. The v0.3.1 legacy snapshot reports it
improved — the runner's signal for a case that passes while annotated.

The issue asked to remove the annotation from
benchmarks/cases/cpp-lsp-parity.yaml. That is not available: the file is
content-addressed by the promotion manifest that binds it, and corrections to
this tier are append-only. So the retirement is recorded as an overlay, which
is what promotion already is.

What this adds

A case entry in a promotion manifest may carry retiredExpectedFailure:

"retiredExpectedFailure": {
  "supersededReason": "<the annotation, verbatim>",
  "evidence": { "file": "...", "sha256": "..." }
}

The validator requires that

• the historical document actually authors the annotation — you cannot retire
  what was never there;
• supersededReason matches it byte-for-byte, so the manifest records what was
  withdrawn rather than deleting it, and cannot drift from the frozen text;
• the manifest carries supersedes, because a retirement is a correction; and
• retirement is one-way — a superseding manifest may retire more
  expectations but may not restore one its predecessor retired. A later
  manifest cannot reinstate an excuse for a case already held to an ordinary
  pass.

Execution staging drops the annotation from its filtered, execution-only copy
of the corpus, so the case scores as an ordinary pass. The historical YAML
stays byte-identical, and legacy-promotion-v1-balanced-core keeps validating
against it — both manifests validate in CI.

The evidence

The retirement rests on the frozen v0.3.1 legacy snapshot rather than a fresh
native run, which is the reference-container confirmation #164 asked for:

execution mode	container, platformScope: canonical_reference
status	improved
queries	2 of 2 successful, 2 of 2 exact-set
usage→declaration	include/parity.h:45:20 — the authored target
false positives / negatives	0 / 0

Scope moved, docs held back

resolve-freeze-scope.sh, run-freeze-candidates.sh, and the reproduction
contract now point at legacy-v2, so the next legacy freeze runs under it.

The docs evidence map deliberately stays on legacy-v1: it describes the
manifest the published v0.3.1 result was frozen under. It moves forward with
the legacy release that supersedes it, not ahead of it.

Verification

• cargo test --locked — 247 passed (5 new), 0 failed
• python3 -m unittest discover tests — 42 tests (2 new), OK
• validate benchmarks/cases — 61 files
• validate-legacy-promotion on both v1 and v2
• validate-legacy-promotion-cohort — N=10, 158 rows
• validate-reproduction-contract.sh — exit 0
• Staging under v1 vs v2 differs by exactly the two annotation lines, and the
  staged corpus still validates as 30 documents / 110 cases

🤖 Generated with Claude Code