[snowykr]

Summary

Closes #17. Builds on #15's completed-reading protection; does not change the Shift-overlap classifier or address #16.

Plain confirmation accepts the reading already displayed, including unfinished input. It no longer means “run language selection again with final context.”

• Keep the displayed engine language for ordinary confirmation; retain final inference for newly typed digit evidence and the existing explicit-correction learning baseline.
• Route Japanese Enter/Space by the visible language.
• Share Japanese preview/confirmation rendering, preserving segmentation and the visible live candidate. Complete only a genuinely pending Japanese n → ん, not an English segment's final n. Learn the visible candidate with its original reading/context metadata.
• Display a standalone letter's final-scored decision immediately, so Tom's does not rely on confirmation to repair Tom'ㄴ. Retain a separate, one-use partial-language seed for the next letter's original hysteresis transition. Force, pin, reset and conversion transitions invalidate it.
• Preserve explicit Esc, Japanese Space/candidate/kana conversion, successful Tab completion, numeric evidence, correction storage and coordinate protections. No new settings, dictionaries, token exceptions or platform ABI changes.

Regressions and intentional behavior

Observed baseline failures: ㄴ → s on confirmation; わたしはg → わたしhag; and a deterministic live-converter fixture replacing 今日はいい天気でs with きょうはいいtenkides. Added tests fail against exact upstream ce7bd934 and pass on this patch.

The matrix covers Enter, Other, failed Tab, direct commit/new-context commit, non-converting punctuation/Space, pending consonants and n, English segments ending n, live conversion off/on, candidate-learning metadata, explicit conversion, Return consumption, forced/pinned edits, first-letter continuation lifecycle and numeric context. FFI coverage also asserts UTF-16 caret positions, surrounding text, consumed boundaries and unavailable surrounding-text access.

The existing ozzy test now explicitly corrects its visible Japanese guess with Esc before Space, rather than depending on a hidden final English decision. That is intentional: a mistaken preview needs explicit correction, not an unseen language switch on confirmation.

The two session golden snapshots are intentionally updated for preview events: 522/674 enko rows and 597/1,349 enkoja rows change events; all 2,023 final texts remain identical. Decision golden snapshots are unchanged. The repository treats these snapshots as binary; git diff --text exposes their event changes.

Local verification completed before opening this PR

On Apple Silicon macOS:

• cargo test --workspace --exclude tacit-tip: 332 passed, one existing ignored subprocess helper.
• cargo test --workspace --exclude tacit-tip --all-features: 472 passed, two existing ignored subprocess helpers (their parent concurrency tests pass).
• Focused core session/Japanese suites: 63 + 70 passed; expanded FFI regression passed in both editions.
• cargo fmt --all -- --check; default and all-feature workspace/all-target Clippy with -D warnings; git diff --check.
• macOS context/controller suites; installer language isolation 8 checks.
• Release Japanese FFI builds for aarch64 and x86_64 macOS; arm64 Swift executable linked; x86_64 Swift typecheck with warnings as errors.
• Lab 16 tests, both report self-tests, Japanese evaluator/Mozc self-tests, bilingual changelog validation.
• Two bounded independent static review lanes approved the final policy/state and Japanese renderer. These reviews are not independent runtime replications.

Windows-only tacit-tip is excluded on macOS as in Unix CI. Windows/Linux/Android runtime, physical keyboard timing and a newly installed IME were not tested. Universal Swift linking is not claimed: the previously established local x86_64 Swift compatibility-library limitation remains; no packaging workaround was added. Platform source/ABI is unchanged, and the existing native controller tests were rerun. Installed app, personal settings/learning and rollback files were not changed.

Accuracy and performance scope

Compared exact upstream baseline with the final candidate, using the existing evaluators and the same datasets:

• --profile enko/enkoja --corpus 3000: 127/156 non-timing model metrics unchanged. This is classifier-level evidence, not proof of session confirmation behavior.
• Actual enko InputSession + FakeTextField switching replay, --corpus 3000 --fold 0/3, covers 7,050 lines across English/Korean sentences, Korean comments and three scenario sets. No previously correct complete sentence became incorrect; all six bucket final-accuracy values are unchanged. One already-incorrect mixed comment changes output (rlac → 김ㅊ, intended 김c); it remains incorrect. This is not a claim of zero per-character changes or representative accuracy.
• Preview tradeoff is explicit: English-sentence wrong-prefix rate 13.682% → 0.769%, Korean-sentence 1.330% → 3.451%, because standalone Latin letters may appear before Korean composition grows. English preservation is reported separately, not hidden in a net score. These are evaluator-defined prefix rates, not real-world user error rates.
• Five alternating runs of the existing no-commit/no-disk session-preview benchmark: zero-history p95 median 31.417 → 31.709 µs enko, 62.875 → 63.875 µs enkoja; run ranges overlap. The first-key path adds a bounded extra inference. These are synthetic preview/completion timings, not end-to-end latency or a speedup claim.

An initial implementation contaminated growing-word hysteresis with the standalone first-key guess. The wider replay caught 10년/소재 regressions despite the focused suites passing. The separate continuation seed and focused lifecycle regressions address that cause; the final case-level comparison above was rerun afterward.

No release or app installation is included.

[helloiamnotth3devnicetomeetyou]

Reviewed the exact change from ce7bd93 to 8b00148. No blocking findings in the standards or issue #17 contract review.

Independently verified on Windows:

• cargo test --workspace --all-features
• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• The focused unfinished-reading and Japanese confirmation regressions fail on the exact upstream baseline and pass on this patch.
• 9,000 additional deterministic synthetic sequences preserve the visible reading on confirmation, except the permitted pending Japanese n -> ん completion.
• Both session snapshots retain all 2,023 final texts; the 522/597 changed event rows are preview changes.

The first-key continuation lifecycle, explicit-correction provenance, shared Japanese rendering, and visible-candidate learning are consistent with the proposed contract. The disclosed first-key Korean preview tradeoff and explicit correction for a mistaken Japanese preview are understood.

Issue #16 remains a separate classifier regression and is not resolved by this PR. Physical-keyboard timing, installed IME behavior, and the submitted accuracy/performance measurements were not independently revalidated.

Thanks for the careful reproductions and for documenting the tradeoffs and evidence limits.