test(core): bring every sieval/core module over the 70% mutation bar

[ethan-scitix]

Type

• refactor — code restructuring, no behavior change

_{(tests only — no source file is touched)}

Summary

The mutation-score gate in sieval/core/CLAUDE.md (≥70% per modified module) has never been runnable — see #68 for the two config bugs. The first sweep it has ever been possible to run covered all 5,381 mutants in sieval/core: overall 76.7%, so the package is broadly healthy, but four modules sat under the bar.

This PR brings all four over it. Tests only; no source file changes.

module	before	after
runners/resume_gate	41.9%	87.8%
tasks/meta	61.4%	92.9%
tasks/anomaly	64.6%	70.8%
tasks/progress	66.7%	71.5%

All 26 sieval/core modules are now at or above the bar — the first time that requirement has actually been met rather than merely stated.

What was actually missing

Not coverage — every one of these modules was already exercised. What was missing were assertions with discriminating power.

resume_gate (the worst, and the one guarding the firmest contract). The repo states it plainly: "Precise reproducibility is a product contract, not a nicety. Safety guards ship strict-only." Yet:

• reason was asserted as != "". The four reject reasons lead an operator to four different fixes, so swapping any two passed every test — including a swap that sends them somewhere that cannot work.
• format_identity_reject_message had no tests at all — 18 mutants with nothing to observe them, and the function was not even imported. It is the message for resuming into a directory another task produced, where the stated stake is handing back someone else's report without running a sample.
• local and dev are independent markers. The run-side test used 0.6.1.dev3+gxyz, which carries both, so it could not tell the or in that guard from an and. A build tagged only local would have fallen through to the series check and resumed.

tasks/meta. task_meta_from_dict carried 84 survivors: the whole deserializer was unverified, so any field could be read from the wrong key, dropped, or swapped with its neighbour. meta/index.json is how every consumer outside this process learns what a task is.

tasks/anomaly. Two counters over the same data (sample_details counts occurrences, rollout_details counts rollouts — a swap misreports how widespread an anomaly is); save's temp-file-then-rename atomicity, since a truncated anomalies.json is worse than none (it looks present, so regeneration is suppressed); and rules_hash pinned to its current value, because editing any rule's prose rotates every stored report fleet-wide and belongs in a diff rather than in production.

tasks/progress. position is how MultiTaskRunner keeps concurrent runners' bars from drawing over each other; collapsing it to a constant fails no behavioural test. Also that _enable_log is exactly "show_progress and not a TTY" — silencing progress must silence both channels, not swap one for the other.

One method note worth keeping

anomaly took four rounds. The first three guessed where the survivors were — read the source, judge which contracts matter, write tests. The last of those rounds killed exactly zero mutants.

Dumping the actual diff of all 134 survivors changed the answer completely: they were in @sieval_detection_rule's own derivation logic, not in the detection or reporting code. And the already-pinned rules_hash could not reach them, because the built-in rules register at import — a test reading the resulting registry sees what the derivation produced once, never the derivation itself. Killing them required registering a rule through the decorator inside the test.

Read the survivors before writing the tests.

Related Issues

Refs #68 — not a hard dependency. The two PRs touch no file in common: this one changes four test files, #68 changes none of them, and these tests pass with or without it. What #68 carries is the fix that makes mutmut runnable at all ([tool.mutmut].also_copy omits sieval/__init__.py, so mutants/sieval is not an importable package and the run dies before any mutant executes). So #68 is needed to reproduce the scores in this PR, not to merge it.

Merging #68 first is still the tidier order — it means the numbers here are verifiable the moment this lands — but either order is safe.

Test Plan

Automated

• Lint/format clean (ruff check && ruff format --check)
• Type check clean (ty check)
• Unit tests pass — 3268 passed, 0 failed (about 90 new)

Manual

• Mutation score measured before and after for each module, scoped per module, numbers in the table above
• No source file modified — git diff --stat is four test files

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring
• No new upper-layer dependencies added to core/
• Deleted code verified — nothing deleted

🤖 Generated with Claude Code