feat: zero-download semantic fallback — algorithmic embeddings + code-aware FTS (#169, Phase B) - #181

Phase B — full #169 on the merged #173 seams

Reimplements the zero-download semantic fallback on the provider / lifecycle /
factory seams from PR #175, with no monolith and the root cause fixed.

Closes #169

1. Fallback provider, selected by the factory per ModelState

• embedding_fallback_provider.AlgorithmicEmbeddingProvider — the SECOND
  EmbeddingProvider (the neural EmbeddingEngine is the first). Deterministic,
  download-free, same EMBEDDING_DIM; delegates to shared.algorithmic_embedding
  (TF + Random Indexing + within-document co-occurrence bridging; sources in the
  module docstring).
• embedding_factory.use_fallback(state) is the total state→provider map:
  LOADED → neural; PACKAGE_ABSENT, MODEL_FILES_ABSENT, LOAD_RAISED,
  UNINITIALIZED → fallback. No state maps to nothing (test asserts totality).
• The lifecycle no longer blocks or crashes: MODEL_FILES_ABSENT / LOAD_RAISED
  serve fallback this session and fetch in the background
  (embedding_downloads.py, honouring CORTEX_EMBEDDING_ZERO_DOWNLOAD), upgrading
  next session.

2. Root-cause fix (the PR #172 empty-stamp bug)

SqliteMemoryStore stamps embedding_model unconditionally on write, outside
the _has_vec branch — so a store built without sqlite-vec still records
'fallback', never ''. Regression test: sqlite-vec absent + sentence-transformers
absent → stamp 'fallback'.

3. Loud engagement + telemetry

One WARNING when the fallback engages; get_telemetry surfaces embedding_mode
(neural / fallback / unknown) via factory.current_embedding_mode.

4. Code-aware FTS on all recall paths

camelCase/snake_case sub-tokens at index time (augment_content) and query
time — recall WRRF _signal_fts, entity MATCH, and
recall_helpers.compute_vector_fts. memories_fts converted external→self-content
with a one-shot rebuild+reindex migration. Test: a lowercase query recalls a
camelCase symbol memory through the recall_helpers path.

5. Scheduled upgrade

consolidation/embedding_upgrade.run_embedding_upgrade_cycle re-embeds
fallback-stamped memories with the neural encoder and restamps them 'neural' once
the model is available; wired into consolidate's compress phase, count in the
response. Test: fallback memories + model available + cycle → restamped neural.

6 & 8. Genuine absence + mixed-store invariant

• Subprocess test: sentence_transformers genuinely absent (import hook raises,
  not monkeypatched) + CORTEX_EMBEDDING_ZERO_DOWNLOAD → remember + recall
  round-trip with zero network, stamp 'fallback', PACKAGE_ABSENT.
• Mixed store: neural and fallback vectors never cross-rank (query-space vec
  filter) — a neural vector identical to a fallback query is excluded.

Benchmark — LongMemEval-S, SQLite three-way (n=50, this code)

Fallback vs no-vector: ΔMRR +0.102, ΔRecall@10 +20.0 pp → fallback beats the
no-vector baseline (adoption criterion met). Results + MANIFEST under
benchmarks/results/semantic-fallback-169/.

Tests / gates

• New: test_semantic_fallback_169.py (13 tests: provider/selection/stamp/mixed
  store/upgrade/subprocess/migration), test_algorithmic_embedding.py,
  test_code_tokenize.py. Existing embedding suite (56) passes unchanged.
• ruff check . + ruff format --check clean. Full-suite tally reported on the PR
  thread once the run completes.
• Vector-path tests carry skipif(not sqlite-vec); they run on the [sqlite]
  CI legs (test-sqlite / test-windows) and skip on the no-[sqlite] matrix.

🤖 Generated with Claude Code