index deadline interrupts nested dedupe savepoint and masks IndexDeadlineExceeded

[EtanHey]

Problem

On an M1 production index run with the default 1,800-second deadline, APSW's progress handler interrupted a dedupe merge inside the outer batch transaction. merge_duplicate_chunk() then attempted ROLLBACK TO dedupe_merge; SQLite reported no such savepoint: dedupe_merge, masking the intended deadline exception. The CLI exited 1 as a generic indexing failure.

Observed terminal evidence:

SQLITE_INTERRUPT ... [/* TRIGGER chunk_tags_update */ UPDATE chunks SET ...]
Error: no such savepoint: dedupe_merge
[ROLLBACK TO dedupe_merge]

The run had made durable incremental progress before the deadline, but operationally reported the misleading savepoint error instead of IndexDeadlineExceeded. At the measured rate (~2 files/minute for 342 files), the default 30-minute bound is also structurally too short for this host.

Likely path

• VectorStore.upsert_chunks_batch() installs the APSW deadline progress handler around an outer transaction.
• merge_duplicate_chunk() opens SAVEPOINT dedupe_merge and catches a generic exception.
• When the progress handler interrupts an update, the cleanup path issues ROLLBACK TO dedupe_merge even though the interrupted transaction/savepoint is no longer available, replacing the primary exception.

Acceptance criteria

• Preserve and surface IndexDeadlineExceeded when the progress handler expires, including through nested dedupe helpers.
• Never mask the primary exception with no such savepoint during rollback/release cleanup.
• Add a regression test that forces the deadline while a dedupe merge is active.
• Make the scheduled runtime bound realistic/configurable for large M1 indexes, or make the command explicitly resumable/incremental with a successful deadline exit contract.
• Emit enough progress/exit telemetry to distinguish a clean bounded run from a genuine indexing defect.