Run mutation testing on the core modules, not by hand

[yuanhao]

Fourth of the trustworthiness items.

Why

Hand-run mutation testing found a decorative test nearly every time it was tried during 0.18 work. Tests that could not fail for the reason they stated:

• EVENT_VARIANT_COUNT — compared two hand-written lists; a new variant appeared in neither (EVENT_VARIANT_COUNT does not guard what its failure message claims #137)
• unpriced_presets_report_unknown_rather_than_free — survived flipping || to &&
• file_backend_evicts_oldest_to_stay_under_its_cap — keys sorted in write order, so the tiebreak agreed with mtime and it passed either way
• A shared-state leak test that survived the function under test being renamed
• assert_eq!(seen.len(), samples.len()) — tautological, in the PR that fixed decorative assertions
• A claude_sonnet_5 price assertion pinning one field of four, which is how a 1.5x error survived 18 releases
• Two assertions using context::message_text, which returns empty for anything but a ToolResult — so one passed having inspected nothing

Seven, found by hand, in one release cycle. The pattern is not going to stop on its own.

Proposal

cargo-mutants over the modules where a silent failure costs the most:

• src/context.rs — compaction, truncation, execution limits
• src/agent_loop.rs — the loop itself
• src/llm_compaction.rs — the state machine
• src/provider/model.rs — pricing

Not per-PR: it is slow, and the signal needs a human to triage. A scheduled run (weekly, or pre-release) with the surviving mutants reported is the right cadence.

Acceptance

• cargo-mutants runs against the listed modules and the output is readable.
• A baseline is recorded — the point is the trend, not a one-off number.
• Known-acceptable survivors are documented so triage is about what changed, not re-litigating the same ones.

Caveat worth stating

Mutation score is a means, not a target. A test suite optimised for killing mutants can still be one that nobody can read. The value here is specifically finding assertions that cannot fail, which is a narrower and more useful signal than coverage.