ci(mutants): expand mutation testing scope to dispatch + classify, enable on PRs

[Destynova2]

Summary

• Adds mutation testing for src/server/dispatch/ (4 shards: mod.rs split 0/2 + 1/2, retry.rs, provider_loop.rs) and src/routing/classify/ (2 shards: mod.rs, classify.rs) to the existing main-only matrix.
• Introduces a PR-triggered mutants-pr job that runs only on the Rust files the PR touches within the curated scope (router, dispatch, classify, dlp). Wall-clocks at 25 min, writes a job summary + sticky PR comment, and stays informational — the full matrix on main remains the source of truth.
• Adds scripts/mutation-pr.sh (diff-based filtering, scope filter, jq summary parsing, GITHUB_OUTPUT wiring).

Why

• The audit flagged that mutation testing only ran on main and only covered router/ + features/dlp/ (5 files). Critical untested modules: server/dispatch/, routing/classify/.
• Running on PRs surfaces regressions before they land on main, where the previous setup just produced post-merge noise.

Scope decisions

Module	Shards	Rationale
src/server/dispatch/mod.rs (575 LoC)	5a, 5b (--shard 0/2 and 1/2)	Largest dispatch file; same split style as pii.rs
src/server/dispatch/retry.rs (478 LoC)	5c	Single shard fits in 60 min budget
src/server/dispatch/provider_loop.rs (275 LoC)	5d	Independent surface, kept separate from retry.rs
src/routing/classify/mod.rs (476 LoC)	6a	Routing entry point, well-tested already
src/routing/classify/classify.rs (512 LoC)	6b	Stateless complexity classifier

resolver.rs, telemetry.rs, inference.rs, tier_match.rs, etc. were not added to the matrix yet to keep total per-PR cost bounded; they can land in a follow-up.

PR sampling strategy (mutants-pr)

1. Diffs BASE...HEAD to find changed src/**/*.rs files.
2. Drops anything outside the curated scope (router / dispatch / classify / dlp). Empty intersection → job skipped (status=skipped-out-of-scope).
3. Runs cargo mutants --file ... -- --lib with timeout --foreground --preserve-status 1500 (25 min cap).
4. Parses mutants.out/outcomes.json via jq for caught/missed/timeout/unviable counts.
5. Writes GITHUB_STEP_SUMMARY + upserts a single PR comment (no spam on re-runs).

Status legend exposed to the PR comment: clean, missed, timed-out, skipped-no-rust, skipped-out-of-scope.

.cargo/mutants.toml

Left unchanged. The existing five exclusions already document each unreachable mutant with a file/line rationale; tightening the regex list is a separate, code-touching effort and was deferred to keep this PR diff scoped to CI infrastructure.

Caveat

If the new mutants-pr job (or the next main run with the expanded matrix) surfaces a wave of surviving mutations on dispatch/ — likely, since the audit found 0 unit tests there — those are tracked as test gaps for the parallel test/dispatch-unit-tests PR. This PR explicitly does not aim to fix them.

This PR could not run cargo-mutants locally to produce concrete survivor counts (the Claude Code sandbox blocks cargo execution). The first run of the workflow on this branch will produce the baseline; the comment it posts becomes the acceptance criteria for the unit-test PR.

Test plan

• CI: mutants-pr job runs (this PR touches .github/workflows/ci.yml + scripts/, both outside the scope filter, so the script should exit skipped-out-of-scope → no mutation work performed).
• CI: existing mutants shards 1-4 still green on the next main push.
• CI: new shards 5a-d, 6a-b run on next main push (informational, continue-on-error: true).
• CI: validate-yaml (actionlint) passes on the new YAML.
• CI: required gate is unchanged — neither mutants nor mutants-pr blocks merge.

🤖 Generated with Claude Code