Prepare v0.2.2 release

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "switchify-prediction"
-version = "0.2.1"
+version = "0.2.2"
 edition = "2024"
 rust-version = "1.97.1"
 license = "MIT"

docs/releases/v0.2.2.md

@@ -0,0 +1,56 @@
+# v0.2.2
+
+Faster whole-word generation, natural generation context, and personal learning
+that is fast, effective and bounded. Worker protocol 2, the model, the tokenizer,
+the bundle manifest and the SQLite formats are unchanged, so existing personal
+databases keep working. The matching v0.2.2 workers carry the faster search.
+
+## Generation (#25)
+
+- **Batched search.** Beam extensions at each depth now run as one batched
+  forward pass. Search stops as soon as further search cannot change the ranked
+  words. For the same context, the words are identical to v0.2.1's on all 1,760
+  frozen queries.
+- **Speed.** Median generation went from 581 ms to 121 ms with the AVX2 worker
+  and from 1,257 ms to 424 ms with the portable worker on a Windows test machine.
+  On macOS ARM CI, with the portable worker, it went from 690 ms to 151 ms in
+  single runs, whose timing varies; one generation in that run failed, which
+  the report does not explain.
+- **Natural context.** Generation now sees the current sentence as typed, with
+  its case and punctuation. On the frozen workload, six-slot top six rose from
+  72.4% to 73.2%. Reranking and statistical prediction still use the
+  normalized context.
+- **Measuring natural context.** `scripts/generation_evaluate.py` gains
+  `--natural-context`.
+
+## Personal learning (#27, #29)
+
+- **Incremental learning.** `Predictor::learn` and `import` add only the new
+  counts unless another connection has changed the personal database. On the
+  production baseline a sentence takes about 8 ms instead of about 190 ms with
+  little learned, rising to about 20 ms after 2,000 learned words.
+- **Weighting.** Learned counts are kept apart from the baseline and combined
+  when predicting. One learned occurrence counts as up to 500 baseline
+  occurrences, less as more is learned, so learned words approach a bounded
+  share instead of outgrowing the baseline.
+- **New and rare words.** A word the baseline lacks, or has seen fewer than
+  three times, counts once it has been learned twice. A single misspelling the
+  baseline lacks or rarely sees is therefore not suggested.
+- **Results.** On a synthetic person, learned words reached the six slots for
+  46.2% of the 78 queries on personal words instead of 1.3%, and top six over
+  all 445 queries rose from 50.3% to 60.4%, with no misspelling suggested.
+  General English accuracy was unchanged with that learning present. These
+  were measured with v0.2.1 generation. See
+  `docs/personal-learning.md` and `scripts/personal_evaluate.py`.
+- **Memory.** The baseline is no longer copied, so peak predictor memory falls
+  from 362 MiB to 134 MiB with or without learning.
+- **Behaviour change for existing users.** Learned counts now weigh much more
+  than the previous 5×, so learned words appear far more often. A word learned
+  only once that the baseline lacks or rarely sees no longer counts until it is
+  used again.
+
+These are regression measurements on synthetic and frozen fixtures, not
+evidence of real users' accuracy. Prepared CI bundles are unsigned; platform
+applications sign their embedded workers through their own packaging
+workflows. Publishing this release, and updating Switchify PC's worker assets
+and library pin to it, need separate approval.

neural/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "switchify-prediction-neural"
-version = "0.2.1"
+version = "0.2.2"
 edition = "2024"
 rust-version = "1.97.1"
 license = "MIT"

neural/worker/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "switchify-smol-worker"
-version = "0.2.1"
+version = "0.2.2"
 edition = "2024"
 rust-version = "1.97.1"
 license = "MIT"