test(assistant): cover the compaction provider view

[juacker]

Implements roadmap item R6.3. Sibling of #204 (which covered the compaction repository functions); this covers the function that turns one of those rows into an actual truncation of what the provider sees.

provider_history_messages_with_compaction had zero tests, while everything around it in compaction.rs is well covered — the size gates, the window selection, the summary clamping. It is the function that decides what actually reaches the model.

Eight tests, one per property: the summary leads the view; the boundary is inclusive; the summary is emitted once from the front rather than left in the tail it always falls inside; the no-compaction path still strips summaries; an orphan summary in the retained tail is dropped; all four unusable rows fall back to the untruncated history; the view carries the stored summary text and marker; the cut happens exactly where the row says, even mid tool group.

Two findings worth the reviewer's attention, both recorded in the test docs rather than acted on:

• The four "unusable row" guards are unreachable today. Both callers get their row from latest_completed_compaction, whose WHERE adds IS NOT NULL on both ids; and ON DELETE SET NULL plus the pool's foreign_keys means a deleted message nulls the column rather than leaving it dangling, which that same filter then excludes. The guards are one half of an invariant whose other half is a WHERE clause in another file, so the test pins the half that lives here.
• compact_session_history writes the summary message before complete_compaction, non-atomically. If the second call fails, the summary is durable while its row stays running forever — nothing writes CompactionStatus::Failed and nothing reaps a stale running row. That is what makes the orphan-summary case reachable, and it is the metadata filter in the tail that saves it. Flagged for a follow-up; no production change here.

Verification. Nine mutations of the function, eight killed, every test killing at least one. The survivor is the message.id != summary_message_id half of the tail filter, unreachable because the one site that writes a summary always sets the marker the other half keys off — documented as redundant defence at the test that would otherwise appear to cover it. An earlier draft of the orphan test was killed by the skip rather than by the filter it named; the sweep caught that and the fixture was re-derived.

A source-only adversarial review found five false claims in the first draft's doc comments (including the reachability story above, which was wrong in three separate ways) and one redundant test, since removed. All applied.

1060 → 1068 tests. cargo fmt clean, cargo clippy --all-targets clean, cargo test --lib 1068/0, bindings.ts untouched. No production code changes.