fix(556): gate schema-echo dicts at the persistence boundary

[Muizzkolapo]

Summary

StorageBackend.write_target now refuses to persist a record whose action namespace is the compiled JSON Schema (the Ollama format shape: {"title","type":"object","properties",...}) instead of LLM output. A resume / carry-forward record reaches write_target without an LLM call, so the in-flight _reject_schema_echo_items guard never runs and the schema dict was written verbatim into target_data.

The new _gate_schema_echo_deltas runs after delta extraction and before _write_target_raw: a namespace matching is_schema_echo is replaced with the make_schema_echo_error sentinel and the record gets a FAILED/PARSE_ERROR disposition, so the failure surfaces on the next read exactly like the live-LLM parse-error path. It lives on the abstract base, so every backend inherits the invariant. Reuses is_schema_echo / make_schema_echo_error / PARSE_ERROR as-is — no new constants, no collector change.

Verification

• RED: 5 of 7 tests fail on main (echo persisted verbatim, no disposition), including a varied-action-name test so a hardcoded-name gate can't pass.
• GREEN: full suite 8126 passed; ruff + mypy clean.
• Review (HIGH tier, 3 blind lenses):
  • Correctness: YES — gate at the right chokepoint (covers force_full + delta branches), guards mirror _extract_delta, replacement reaches _write_target_raw, NotImplementedError caught.
  • Blast radius: YES — sqlite RLock is re-entrant (calling set_disposition from inside write_target is deadlock-safe); every write_target caller checked for false-positives (none carry title+type:object+properties); no SUCCESS-write race.
  • Tests/anti-gaming: flagged that all tests used one action name → fixed by adding test_gate_fires_for_any_action_name (recorded in RED). Real SQLite round-trip, non-tautological assertions, collector untouched.
• Smoke: skipped — scan-project canary divergence is pre-existing (baseline.sh confirms it fails on main too; stale project action registry, unrelated to storage). The gate is a no-op on clean data; the real write_target→_write_target_raw→_read_target_raw round trip + disposition writes are covered by the 7 unit tests.

Deviations from the spec (verify-first)

• set_disposition clears priors. The spec assumed a prior SUCCESS and the gate's FAILED would coexist (UNIQUE(action,record,disposition)). Current set_disposition (sqlite_backend.py) DELETEs all prior dispositions for (action, record) first, so the gate's FAILED replaces a stale SUCCESS — the record ends unambiguously failed (better). Test asserts the real behavior.
• Test path is tests/unit/storage/ (spec said tests/storage/).
• Real-DB before-count is now 0 (spec cited 11 rows on 2026-06-17). Those rows were cleaned since; the gate is a forward-looking invariant — RED proves write_target still persists a schema-echo namespace verbatim on main, i.e. the hole is open.

Out of scope (follow-up)

The upstream planter that stamps compile_unified_schema output as record content on the resume/checkpoint path — suspect sites processing/strategies/online_llm.py:_checkpoint_record and llm/batch/services/processing_recovery.py. This gate is the safety net regardless of which path is wrong.

[Muizzkolapo]

Real-project verification (added)

Ran the gate against a copy of the actual qanalabs code_centered_quiz.db (real SQLiteBackend, real write_target chokepoint — not an in-memory fixture). Wrote a synthetic schema-echo record for explain_distractor_1:

[explain_distractor_1] Schema-echo in delta for source_guid=sg-demo (keys: [...]) — replacing with parse-error sentinel
persisted namespace keys: ['_parse_error', 'raw_response']
  -> schema echo gone?  True
  -> parse-error sentinel present?  True
disposition rows: [('failed', 'parse_error')]

Confirms the gate replaces the namespace and records FAILED/PARSE_ERROR end-to-end against the real project DB schema. Run on a copy; the project DB was not mutated.

Note: the spec's original before-check (11 rows on 2026-06-17) now returns 0 on the live DB — those rows were cleaned since. The gate is a forward-looking invariant (RED proves write_target still persists a schema-echo namespace verbatim on main).