test(mcp-skill): pin declared-field alias-bomb retype vector

[bug-ops]

Summary

• Adds regression coverage for the second alias-bomb resistance vector on RawFrontmatter: retyping an already-declared field (e.g. description) from Option<String> to a buffering type, distinct from the flatten/unknown-key vector covered by the prior regression test.
• Two canary tests pin serde_norway's own buffering behavior directly: a Value-typed retype (compile-blocked in production today, kept as a mechanism baseline) and a deserialize_with retype (not compile-blocked — the realistic silent-regression case).
• A fourth, production-coupled test confirms today's real extract_skill_metadata + RawFrontmatter does not hit the repetition-limit path, and will fail loudly if RawFrontmatter is ever retyped to a buffering shape.
• Adds an ADR-341 addendum (§10) recording the vector and correcting the attribution of its timing figures (issue add regression coverage for RawFrontmatter field-retype alias-bomb vector #359, not the ADR's own §3.3 measurement, which is explicitly scoped to the undeclared-key/flatten vector).
• CHANGELOG entry under [Unreleased] > Testing.

Went through two rounds of adversarial critique (first verdict: significant — no coupling to production code, covered the compile-blocked variant instead of the silent one, an overly loose is_err() assertion; second verdict: minor — wording/doc-order nits plus adopting the production-coupled 4th test). All findings addressed and re-verified.

Closes #359

Test plan

• cargo nextest run -p mcp-execution-skill -E 'test(alias_bomb)' — 4/4 pass
• cargo nextest run --workspace --all-features --lib --bins — 987/987 pass
• cargo test --doc --all-features --workspace — 40/40 pass
• cargo +nightly fmt --check — clean
• cargo clippy --all-targets --all-features --workspace -- -D warnings — clean