test(tasks): cover the agent-task store's claim, sweep and archive logic

[acamarata]

src/tasks/storage.rs (940 lines) had zero tests. Adds 16 covering add_task's defaults, claim_task's exclusivity and re-claim window, update_status's completion-notes guard, interrupt_stale_tasks, archive_done_tasks and prune_activity_log.

Time without sleeping

now_ts() is wall-clock, so the tests write last_heartbeat / completed_at directly with SQL. A sleep-based test for a 90-second reclaim window would either take 90 seconds or be a lie.

The reclaim-window tests straddle the boundary deliberately. The guard is elapsed > reclaim_window * 2:

window	elapsed	expected
default 90s (→180s)	400s	rejected
default 90s (→180s)	100s	accepted
explicit 10s (→20s)	100s	rejected

The same 100s flipping verdict with the timeout is what pins the parameter as actually used — killing both unwrap_or(90) → a constant and * 2 → * 1.

A real gap the spot-check caught

prune_activity_log's fixture originally held entries at 3 days and 1 hour with retention=2 days. The * 86400 → * 3600 mutant moves the cutoff from 2 days to 2 hours — and deletes exactly the same two rows. The test passed while killing nothing.

Fixed by adding a 1-day-old entry, the only value that separates the two cutoffs, and asserting the survivors by id rather than by count. That mutant now dies.

This is the whole reason each suite is verified by hand-applying mutants rather than trusting a green run.

Verification — 7/7 killed

Mutant	Result
reclaim window * 2 → * 1	KILLED
unwrap_or(90) → (0)	KILLED
done-notes guard removed	KILLED
archive cutoff * 3600 → * 1	KILLED
archive's status='done' filter dropped	KILLED
prune * 86400 → * 3600	KILLED (after the fix above)
interrupt_stale_tasks → empty vec	KILLED

cargo test --lib tasks::storage::tests passes; clippy --all-targets --all-features -- -D warnings and cargo fmt --check both exit 0.

Noted, not changed

claim_task's lease branch matches status = 'claimed', and the conflict query matches 'claimed'/'active' — but the schema's CHECK constraint permits only pending/in_progress/in_cr/in_qa/blocked/interrupted/done/deferred. Those predicates can never match. Left alone: narrowing live SQL is a behaviour change, not test coverage. Worth a look separately.