test(assistant): cover the compaction repository functions

[juacker]

Implements R6.2. Test-only: one file, +490, no production code changed.

What was missing

create_compaction, complete_compaction, get_compaction and latest_completed_compaction (repository.rs:1171-1295) had zero tests. The blocker was the hand-written fixture, which never created assistant_compactions; db::test_support::workspace_pool from #200 builds the real schema from the migrations, so the table is simply there now.

What is actually load-bearing here

latest_completed_compaction selects the row that decides which messages the provider is allowed to see. It is not the last line of defence — I initially wrote that it was, and the review disproved it: the one consumer, provider_history_messages_with_compaction (compaction.rs:320-340), re-checks the same two nullable fields and falls back to the unfiltered history in every failure case. It is the first line, and none of its four WHERE/ORDER BY terms had a test.

Each now does. A compaction that is still running, that has no summary message, or that has no window end must stay invisible; and the row that wins is the last one completed, not the last one created.

Two of these states are reachable from the UI

summary_message_id and source_to_message_id are both ON DELETE SET NULL, and delete_message is called from both assistant drivers (engine.rs:2254, local_agent.rs:2227). So deleting the last compacted message — or the summary itself — silently un-compacts the session: the compaction row survives, the summary message may survive, and the next request carries the full history again.

The tests build those two states through delete_message rather than raw SQL. That makes the comments true rather than hypothetical, and pins two of the table's three FK actions for free; the third, the session cascade, gets its own test.

Mutation table

Re-derived from thirteen runs after review, not asserted. Each mutation fails exactly the named test and nothing else.

Mutation	Killed by
status = '"completed"' dropped	ignores_a_running_one
summary_message_id IS NOT NULL dropped	ignores_one_without_a_summary
source_to_message_id IS NOT NULL dropped	ignores_one_without_a_window_end + deleting_the_window_end_message_…
ordering flipped to ASC	picks_the_newest_completion
ordering by created_at	picks_the_newest_completion
error = NULL dropped	complete_compaction_sets_the_summary…
touch_session dropped, create side	create_compaction_touches_the_session
touch_session dropped, complete side	complete_compaction_sets_the_summary…
protocol/model binds swapped	round_trips_every_field
summary/run-id binds swapped	round_trips_every_field
session ON DELETE CASCADE dropped	deleting_a_session_deletes_its_compactions
window-end ON DELETE SET NULL dropped	deleting_the_window_end_message_…
summary ON DELETE SET NULL dropped	ignores_one_without_a_summary

Review

Delegated adversarial pre-PR review against d25e0c8: shape ship_with_changes, correctness rework. Three must-fixes, all mine, all applied:

1. ignores_a_running_one could not fail with the status filter deleted — the row also had a NULL summary, so a second guard was doing the work. My own first mutation sweep never mutated that term, so I claimed a kill I had not observed. The row now carries a summary, making status the only excluding term, and the whole table was re-derived rather than patched.
2. The claim that renaming CompactionStatus::Completed would silently make every compaction invisible was wrong: the CHECK constraint rejects the next completion and map_compaction_row stops parsing old rows — both loud. The assertion stays (it moves a runtime failure to the rename site) with an honest justification.
3. "A state no code path writes today" was false, per the ON DELETE SET NULL above.

All five should-fixes applied too: a weak created_at/completed_at discrimination in the ordering test (both timestamps are now forced, with no ties, pointing the two orders in opposite directions), the speculative production story behind completion-ordering, created_run_id never round-tripped non-NULL despite being the dangerous adjacent bind, complete_compaction's second touch_session untested, and the invented failed state. Nits folded in.

Follow-ups this found and does not close

• provider_history_messages_with_compaction — the pure function where a compaction row becomes an actual truncation — still has no tests. It takes a slice and is cheaper to test than anything here. This is the next item.
• Nothing in the tree ever constructs CompactionStatus::Failed. A compaction whose summarize_window fails stays running forever, across restarts, and recover_stale_runs sweeps only assistant_runs/assistant_tool_calls. No fail_compaction, no stale-compaction recovery.

Gates

cargo fmt --check clean · cargo clippy --all-targets -- -D warnings clean (under #202's new policy) · cargo test --lib 1060 → 1074, 0 failed, verified by name-list diff — 14 added, none removed · src/generated/bindings.ts untouched.