fix: datetime round-trip in _insert_cb + clearer Rust callback error (closes #36)

[ashish-jabble]

Summary

Closes #36. Two scoped commits:

1. fix(python): parse timestamp/shipped_at back to datetime in _insert_cb — the actual bug. Row crosses Python → JSON → Rust queue → Python _insert_cb on its way through the drainer. After json.loads, timestamp / shipped_at are ISO strings; both PostgresStore.insert and SQLiteStore.insert call _utc_iso(row.timestamp) which reads dt.tzinfo and crashes with AttributeError on a str. The bare except Exception: return 1 then swallows the cause, the drainer dead-letters the row after retry-exhaustion, and the only on-disk record of the audit event is the stdout NDJSON.

   Fix reconstructs the two datetime fields from their ISO strings before constructing AuditRow. Also replaces the bare except with logger.exception so any future drainer-side failure surfaces in the fasten logger instead of being silently masked.

2. fix(core): report CallbackFailed instead of InvalidTableName for non-zero callback rc — the diagnostic-misdirection issue. When the host-language _insert_cb returned non-zero, fasten_core was wrapping it as Error::InvalidTableName(format!("insert callback rc={rc}")). The drainer's audit_drain_failed log then rendered:

   invalid table name "insert callback rc=1":
     must match ^[A-Za-z_][A-Za-z0-9_]*(\\.[A-Za-z_][A-Za-z0-9_]*)?$
   

   — an SQL-identifier validation message run against a callback return code, which sent me ~30 min down a "what table name is invalid?" rabbit hole before realising. New Error::CallbackFailed(i32) variant carries the rc unambiguously: insert callback returned non-zero rc=1. Maps to FastenErrorCode::ErrBackend (the callback's destination is a store backend; same failure category).

Tests

• Python — new regression in python/tests/test_engine_init.py exercises the queue-mode drainer end-to-end and asserts the store receives a datetime, not a str. Fails on main, passes on this branch. Full suite: 229 of 229 lib tests pass (the 17 store_compat parametrised cases skip without a configured Postgres DSN — unrelated).

• Rust — 2 new tests in fasten-core/src/ffi/mod.rs::tests:

  • callback_store_failure_reports_callback_failed_variant — non-zero rc produces CallbackFailed(7), human-readable form is "insert callback returned non-zero rc=7".
  • callback_store_success_returns_ok — rc=0 returns Ok(()).

  cargo test --release --features all passes: 12 drainer_conformance + 52 lib + 2 new = 66, no regressions.

• End-to-end — verified in a downstream adopter (fromSpec factory pinned at this branch via git URL). emit → flush → SELECT FROM fasten.audit_log returns the row instead of dead-lettering.

Compatibility

Error::InvalidTableName is unchanged — validate.rs, store/sqlite.rs:417 assertion, store/postgres.rs:493 assertion, rust/src/engine.rs:104, and the Go / Swift binding tests all continue to use it for the cases where it's actually correct (table-name regex rejection). The drainer_conformance tests' BrokenStore / FlakyStore still construct InvalidTableName("store_down") as a stand-in store-failure shape — left alone since those tests assert drainer behaviour, not the variant kind, and changing them would obscure the narrow scope of this fix.

[praveen-garg]

lgtm!