Add next-word prediction core to the candidate list

[mkpoli]

Why

The IME's autocompletion feels like it "only completes one word" — because it literally does. CandidateList::build only completes the current partial word; the moment a word is committed the suggestion engine goes dormant (key_event_sink.rs:100 gates on a non-empty buffer, candidates.rs:46 returns an empty list for an empty word). It never predicts the next word, even though the n-gram engine already can (Suggestions::predict_scores, next_words, default_words).

This is the first of two PRs adding next-word prediction (azooKey-style predictive continuation). Note: for Ainu the "conversion" problem that drives Japanese IMEs (kanji homophones) barely applies — output is kana/latin — so the real win is predictive continuation, which is exactly what's missing.

What

Adds CandidateList::predictions(prev2, prev1, suggest, max):

• Ranks candidates by the blended trigram+bigram context scores (predict_scores), best first, alphabetical tie-break for a stable order.
• Fills remaining slots with the globally most-frequent words so the popup is useful at a cold start (unknown/sparse context).
• Duplicate-free, capped at max. Unlike build, there is no "typed word" at index 0 — every entry is a real predicted word.

Pure, host-independent logic. Four new unit tests (context-first ranking, trigram-beats-bigram, cold-start frequency fallback, max/dedup) — all green, and cargo clippy -D warnings clean on the MSVC target.

Scope

This PR is logic only — nothing yet consumes predictions(). The TSF integration (a no-composition predictive popup, accept-and-insert, and the key gating that surfaces it after a commit) lands in a follow-up PR, which will need Windows verification since TSF behavior can't be exercised on Linux.

Summary by CodeRabbit

• New Features
  • Added smarter next-word suggestions that prioritize context when available, fall back to common words when not, and avoid duplicate entries.
  • Suggestions are now consistently ordered, with the most relevant matches shown first and ties broken predictably.
• Tests
  • Added coverage for context-aware ranking, cold-start suggestions, result limits, and duplicate-free behavior.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 0e475de9-7128-4e61-abd1-caa08ef3c310

📥 Commits

Reviewing files that changed from the base of the PR and between 26561ac and 01678f6.

📒 Files selected for processing (1)

• src/candidates.rs

---

📝 Walkthrough

Walkthrough

Adds a new CandidateList::predictions constructor in src/candidates.rs that builds a duplicate-free, capped list of next-word predictions, ranking context-based scores first and falling back to default frequency-based words. Adds corresponding unit tests.

Changes

Next-word Prediction Candidate List

Layer / File(s)

Summary

Predictions constructor and tests src/candidates.rs

Adds CandidateList::predictions(prev2, prev1, suggest, max), which ranks context-predicted words first using suggest.predict_scores, sorts by descending score with alphabetical tie-break, fills remaining slots with suggest.default_words(max) deduplicated via HashSet, caps at max, and initializes selection index to 0; adds tests for context-first ranking, trigram-over-bigram precedence, cold-start frequency fallback, and max/dedup enforcement.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: adding next-word prediction support to the candidate list.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/next-word-prediction-core

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}