Activity

1. buildingjoshbetter

   added

   bugSomething isn't workingSomething isn't working

   source:auditFound by code auditFound by code audit

   P1Correctness bugs with realistic triggers; perf felt every sessionCorrectness bugs with realistic triggers; perf felt every session

   perfPerformancePerformance

   on Oct 8, 2026
2. buildingjoshbetter commented on Oct 8, 2026

   

   buildingjoshbetter

   on Oct 8, 2026

   OwnerAuthor

   More actions

   Follow-up: exact source tracking on legacy/custom schemas

   Evidence class: reproduced with synthetic in-memory SQLite on combined draft commit 4fdcb50fa056f17eac3ce6c792a5e798b8be890d. This is a draft-foundation correctness gap; it is not evidence that the new maintenance runner is deployed. Related FTS scope: #751. Standalone rebuild draft #802 has a separate bridge for these cases; that bridge does not certify canonical maintenance or FTS.

   Reproduction A: collated source edits

   1. Load the actual truememory/storage.py module with runpy.run_path; create an in-memory connection.
   2. Execute its _SCHEMA_SQL, replacing only content TEXT NOT NULL with content TEXT COLLATE NOCASE NOT NULL.
   3. Call _initialize_maintenance_tracking(conn), insert synthetic-alpha, and commit.
   4. Record revision, correction_count; update content to SYNTHETIC-ALPHA and commit.
   5. Compare the source row, maintenance state and stored FTS content.

   Observed: source content changes, but maintenance remains (1, 0) and FTS retains the old content. Repeating with RTRIM and a trailing-space edit produces the same result. The ordinary BINARY control advances to (2, 1) and updates FTS.

   Cause: trigger predicates use old.field IS NOT new.field, so column collation can hide a byte-level edit. This can incorrectly retain successful provenance and delay required recomputation. FTS token membership may be unchanged for these specific case/space edits; the stored FTS content is nevertheless stale.

   Reproduction B: replacement removes an older source row

   1. Use the same actual schema but declare content TEXT NOT NULL UNIQUE.
   2. Leave PRAGMA recursive_triggers=0, install source tracking, then insert (id=1, content='synthetic-alpha') and (id=2, content='synthetic-beta'); commit.
   3. Run INSERT OR REPLACE INTO messages(id,content) VALUES(3,'synthetic-alpha'); commit.
   4. Inspect source IDs, FTS row IDs and source revision counters.

   Observed before counters (revision, inserts, corrections, high_id, nonappend_revision): (2, 2, 0, 2, 0). After: (3, 3, 0, 3, 0). Source IDs are [2, 3]; FTS IDs are [1, 2, 3]. The counter token incorrectly certifies an append-only transition while source ID 1 has disappeared. The missing delete-trigger notification is relevant to custom unique indexes and indexed primary-key forms too; those variants still require dedicated controls in this canonical path.

   Implementation plan

   • Compare each tracked field by both SQLite storage type and byte representation, with null-safe comparisons. Preserve no-op behavior for truly unchanged fields and derived-only writes.
   • Bind the canonical tracking contract to schema/index identity or explicitly reject unsupported schemas before issuing a freshness token. For unique constraints that permit an implicit REPLACE deletion, conservatively classify writes as corrections unless deletion visibility is independently guaranteed.
   • Repair FTS consistency using an explicit supported transaction contract. A local pragma alone does not cover writers on other connections. Avoid a full-table cleanup on every ordinary append; measure any custom-schema fallback before choosing it.
   • Version trigger migration and rotate the tracking epoch when old semantics leave an uncertified interval. Preserve caller transaction/savepoint ownership, previous good derived rows and pending user work on failure.
   • Keep the standalone rebuild bridge independent until equivalence is proven; do not broaden its existing claims.

   Acceptance gates

   • Actual SQLite controls for NOCASE, RTRIM, byte-identical no-ops, null/type changes, default schema, custom unique indexes and indexed primary keys.
   • INSERT OR REPLACE and UPDATE OR REPLACE with recursive triggers both off and on, including a second writer connection.
   • Source token invalidation whenever any previously covered source row changes or disappears; no false append-only certification.
   • FTS row/content parity and exact result checks after replacement, rollback and retry.
   • Old schema/trigger migration invalidates only untrusted provenance; migration failure preserves caller data and old valid output.
   • Default-schema append/episode benchmarks retain their recorded semantics and show the added bookkeeping cost.

   All inputs above are independently generated strings. No personal database, memory, transcript or machine identifier was read for these reproductions.
3. buildingjoshbetter commented on Oct 8, 2026

   

   buildingjoshbetter

   on Oct 8, 2026

   OwnerAuthor

   More actions

   Engine-routing integration blocker

   The unpublished engine-routing candidate produced 13 failures, 2,662 passes and 1,195 passing subtests in the full synthetic GPUBox suite with sqlite-vec enabled. Log SHA256: 4def1dea951493035ab915d9aa368d8bf2e54e1dbcba18bbcef3854cde49cb8b.

   Independent event-driven reproduction confirmed a real foreground regression. Initial automatic clustering holds the model state lock while capturing source/vector data. A valid add can finish encoding, acquire SQLite writer ownership, then fail the nonblocking model publication fence. The source row rolls back. Increasing the scheduling threshold only postpones this race.

   The correction is a separate prerequisite: acquire foreground model admission before opening an owned SQLite write transaction, retain that admission through commit, and preserve nonblocking behavior for caller-owned transactions. Borrowed model capture must use an already-loaded protected model or return typed deferral; it must not load a model or wait for the model lock while holding the caller's writer. Genuine generation changes remain rejection conditions.

   The direct rebuild tests also need explicit automatic-maintenance isolation because they test exclusive rebuild ownership. Their batching and durability assertions will remain intact. Fixture isolation alone is not accepted as the production fix.

   The scheduler candidate remains uncommitted and unpublished. Deterministic regressions, independent review and a clean full suite are required before publishing it. No native performance or MPS result is claimed.