CI canary root cause: get_shard() returned a raw sqlite3 connection —
virtual tables need the extension loaded PER CONNECTION, so recall() on a
shard crashed 'no such table: entries_vec' in clean environments. Also:
create_shard's schema lacked entries_vec entirely (main store had it,
shards didn't — recall routes to shards). Both fixed; 3 red-first tests
frozen green (canary failure class, schema completeness, shard reuse).
Gate finding 8: the canary's trivial-task check ran on an EMPTY store —
the capsule-level zero-memory proof was vacuous. It now seeds a
high-importance fact distractor plus a personality row before
assembling; the trivial task must still produce zero memory blocks and
zero memory tokens.
Also lands the discharge record: docs updated (floor-in-every-mode
claim, known ranking bias, typed persona lane, write-time memory_type,
migration-time backfill) and the plan's Completion Record gains the
Discharge Record section mapping every gate finding to its commit and
red-first evidence.
Gate finding 7 follow-up: gate_consolidated itself is deleted (it rode
into the finding-1 commit with the memory_store rewrite); this commit
fixes canary_query_memory's gate check, which inspected whichever
memory_store a sys.path shim resolved first (the DEPLOYED scripts copy,
in practice) — it now proves the retirement deterministically via an
AST parse of the repo copy. Suite and canary battery stay green.
Gate findings 5+6 (test-only):
- finding 5: the operator-level tests passed trivially on empty tmp
stores ('block is None or ...'). They now seed the store and assert
the positive branch — a real selector-sourced block with exact source,
and the history test seeds a completed past task through the canonical
ledger so the priority/label assertions are positive.
- finding 6: test_cached_recall_never_returns_flat_eight_tenths escaped
via 'or source==consolidated'; it now asserts the calibrated recall
shape (rrf/confidence/id present, relevance != 0.8) with shard routing
disabled (it silently rerouted to an empty shard, vacating the old
assertion), plus abstention + positive twins. The importance-fallback
grep now covers ALL fallback slice patterns plus a behavioral twin
(off-trigger high-importance neuron does not inject through the
operator).
Gate finding 4: ensure_memory_type_column ran an unindexed
UPDATE ... WHERE kind='neuron' scan on EVERY connect, and the selector
opens a fresh connection per turn — pure per-turn overhead. connect()
now invokes the migration only when the column is actually missing
(the ALTER stays idempotent); explicit ensure_memory_type_column()
keeps the frozen self-healing semantics (re-classifies unclassified
neuron rows, never touches explicit classifications).
Red-first evidence: a NULL marker row planted on an already-migrated
store was flipped back to 'associative' by a second connect — now the
per-connect scan is gone while explicit ensure() still self-heals.
Gate finding 2: the typed persona source was dead in production — the
operator's single memory contributor calls select_for_task WITHOUT
types, so _persona_recall always returned []; and no production writer
ever set memory_type. The store source now always runs a SECOND typed
query (self/personality) merged into the candidates — still thresholded
inside the store's gate, deduplicated by entry id, may abstain (a
trivial task still gets zero persona rows). add_entry gains an additive
memory_type param (default None = unclassified, eligible everywhere)
plus a --memory-type store CLI flag wired through the dedupe gate;
wired where the writer genuinely knows the type: warm-router neuron
writer -> associative, sleep-time growth sync -> personality
(persona-targeted) / semantic.
Gate finding 3: _persona_recall ignored the selector's pinned db_path
(bare memory_store.connect() opened the module-default store). The pin
is now bound into sources that declare db_path at registration time;
both store-backed sources honor it.
Red-first evidence: no typed call in the live contributor path;
add_entry TypeError; persona block missing from the pinned-only store.
Gate finding 1: min_score was only enforced in recall's dense-only
branch — lexical-only and hybrid paths returned below-floor store rows,
and the selector's store_scored exemption then injected them. The floor
now applies in EVERY mode inside the store's calibrated gate (lexical
best-relevance check additive to the conjunctive FTS gate; hybrid gates
on the dense cosine, the scale min_score was calibrated against); the
selector docstring claim was rewritten to match reality.
Also documents the known ranking-scale mix (finding 9): lexical store
~4-6 vs dense/hybrid RRF ~0-0.13 vs self-scored 0-1 ranked raw together
— abstention unaffected, per-source rank normalization deliberately
deferred (renormalizing the calibrated scales risked false abstentions).
Red-first evidence: lexical-only rows returned at floor 5.5; hybrid
cos~0.37 row injected at floor 0.42 — both abstain after the fix, with
above-floor positive controls.
Nick