Evaluate Taskmaster-2 data mixtures for word prediction

[enaboapps]

Result

Adding 21,000 Taskmaster-2 USER sentences at 1x and 3x did not pass the frozen development criteria. Mean early top-five accuracy changed by -0.128 and -0.751 percentage points respectively. The lighter candidate was evaluated diagnostically on AAC, general English, conversation and a filtered COMM2 set. At two graphemes, top-five accuracy fell on all four sets, although some top-one and longer-prefix results improved.

The 1x database grew by 13.9%; worst warm p95 was 6.885 ms against 6.425 ms for the baseline. Keep the shipping model unchanged.

Change

Adds a reproducible experiment command, pinned source checksums, a predeclared selection protocol, synthetic tests, aggregate JSON results and a report. COMM2 stays out of training; exact overlap checks preserve all existing development/test partitions. Prior legacy-backfill accuracy is included only for matching test hashes, with no cross-run timing comparison.

Production APIs, model files, packaging and Switchify PC are unchanged. No corpus text or experimental databases are committed or published.

Validation

• Completed 17 local scoring runs, each with more than 1,000 queries at prefixes 0 through 4; no personal learning.
• cargo fmt --all -- --check
• cargo clippy --locked --all-targets -- -D warnings
• cargo test --locked: 22 tests passed
• Python unittest discovery: 15 tests passed
• cargo audit --deny warnings
• Verified unchanged production training and baseline hashes; prior newer-mode test accuracy exactly reproduced.
• Independent read-only review of final head 5d595cf: no actionable findings. All required CI passed on this head: https://github.com/switchifyapp/switchify-prediction/actions/runs/37123756202 .

See docs/taskmaster2-results.md for gains, regressions, costs and limitations, and docs/taskmaster2-protocol.md for the reproduction command.

Closes #11.