fix: cascade-clean orphaned vec_episodes rows on import_from_dict force-overwrite

[kohai-ut]

TL;DR

vec_episodes is a sqlite-vec virtual table keyed by episodic_memory.rowid (AUTOINCREMENT). When BeamMemory.import_from_dict(force=True) DELETEs an existing episodic_memory row to replace it, the new row gets a fresh rowid via cursor.lastrowid — leaving the old row's vector embedding stranded in vec_episodes pointing at a rowid that AUTOINCREMENT will never re-issue. Operators with high import churn accumulate dead vec_episodes entries indefinitely.

This PR adds a cascade DELETE FROM vec_episodes WHERE rowid = existing["rowid"] BEFORE the episodic_memory DELETE, plus broadens the cleanup exception catch to sqlite3.Error so a cleanup failure doesn't abort the import mid-loop.

5 regression tests, all passing. Originally flagged by Codex adversarial on PR #84 as a deferred follow-up.

Why this matters

Storage hygiene: long-running deployments that do regular import/export cycles (backups, multi-source sync) accumulate orphaned vec_episodes rows that never get cleaned. Each entry is ~768 bytes (float32) or ~48 bytes (binary quantized) — small per row but unbounded growth at scale.

This is the only DELETE FROM episodic_memory site in production code (verified via grep). Mnemosyne.forget() doesn't currently cascade to episodic_memory at all — that's the separate C17 concern in the ledger.

What this PR does

Single commit:

• import_from_dict cascade-cleans vec_episodes before the episodic_memory DELETE on the force-overwrite path
• Guarded by _vec_available(self.conn) — sqlite-vec is optional
• Hoisted vec_ok = _vec_available(self.conn) from before the embeddings-import loop to before the episodic_memory loop — both code paths now share one check
• Broad except sqlite3.Error (not narrow OperationalError) catches all sqlite3 exception subclasses with WARNING log — working_memory was already committed at line 3978, so propagating a cleanup error would abort the import mid-loop with partial state

/review army findings (all addressed)

Finding	Sources	Fix
Initial "best-effort failure" test dropped vec_episodes before calling import, causing vec_ok=False short-circuit instead of exercising try/except	2-source (Codex structured P2 GATE FAIL + Claude C1 CRITICAL)	Test now monkey-patches _vec_available to return True after dropping the table, so the cascade runs and the DELETE fails, exercising the try/except
Missing end-to-end test with non-empty episodic_embeddings payload — every other test used [], leaving the old_to_new_rowid mapping path uncovered for the force-overwrite case	Claude H1	Added test_import_from_dict_force_with_new_embeddings_no_orphan
except sqlite3.OperationalError too narrow — other sqlite3 exception subclasses propagate and abort import mid-loop	Claude H2	Broadened to except sqlite3.Error as cleanup_exc with WARNING log

What is NOT in this PR (intentional, documented)

• Mnemosyne.forget() doesn't cascade to episodic_memory at all (C17 in ledger) — pre-existing different concern
• memory_embeddings fallback table gets stale content on force-overwrite — different orphan shape, separate concern
• binary_vector column NULL after force-INSERT — pre-existing, not affected by this PR
• Concurrent import_from_dict race — import_from_dict is not advertised as concurrent-safe; out of scope

Test plan

• All 5 tests in tests/test_orphan_vec_episodes_cleanup.py pass
• All 20 tests in tests/test_integration.py pass
• Targeted test_beam.py -k "import or export" run: green
• Codex structured /review: GATE PASS after commit 1's fix
• Claude adversarial /review: CRITICAL + 2 HIGH (all addressed)
• CI: full suite green on Python 3.9 / 3.10 / 3.11 / 3.12

🤖 Generated with Claude Code