[enaboapps]

Closes #27

Predictor::learn and import used to clone the whole baseline and reload every personal count on each call. Now, when PRAGMA data_version shows that no other connection has committed since this predictor's snapshot, they add only the new counts to the in-memory snapshot, with the same 5× weight. Otherwise they reload and validate everything, as before. The fast path also requires the personal database to have exactly the schema initialize creates. Triggers, replacing constraints, foreign-key actions and other changes can give the connection's own write side effects that only the full reload validates. A failed commit clears the recorded version, so the next write reloads. A connection's own commits leave its data_version unchanged. The version is read inside the transaction, so no fallible work follows the commit. refresh_personal and reset_personal record the version too. The schema, API and prediction maths are unchanged.

Results

Production en-aac-oanc-v1 baseline, release build, 50 sentences (ignored test learning_cost_on_production_baseline):

	Median	Max
Before	191 ms	624 ms
After	8.4 ms	11 ms

Validation

• incremental_learning_matches_a_full_reload: after each learned text, the ranked words and score bits match a freshly opened predictor.
• learning_after_another_writer_includes_its_counts: a second writer's counts, and a reset by another instance, are picked up through the full-reload path. Results match a reload throughout.
• A unit test shows the fast path runs for this connection's own writes and its own reset. It also shows that another writer, or a non-standard schema, forces a reload.
• a_trigger_side_effect_is_still_validated_before_commit and a_replacing_constraint_is_still_validated_before_commit check that both schema-based cases are still refused and rolled back, as on main.
• duplicate_imports_and_empty_text_match_a_full_reload.
• Existing concurrency, import-dedup, reset and corruption tests pass unchanged.
• Three independent reviews. The final one found no issues. It confirmed that personal databases created by v0.1.0–v0.2.1 match the standard schema, and it fuzzed word normalization (300k strings).
• cargo fmt --check, cargo clippy --locked --all-targets -D warnings and cargo test --locked pass.

This is independent of #26. Switchify PC needs this before it can learn while someone types.

🤖 Generated with Claude Code

[enaboapps]

Three independent reviews; the last found no issues. All 15 CI checks pass. #30 is stacked on this branch, so don't delete claude/incremental-learning on merge, or #30 will close.