Additive and non-breaking. No existing API changes, no behaviour changes, no change to adviseInDrive or to any threshold; English remains the default. The enums remain the API — if you already localize them yourself, nothing here is imposed on you.

Japanese safety-word strength — corrected (read this if you render ja) #

Two ja strings were materially WEAKER than the English they translate. A mis-calibrated Japanese safety word is worse than English: the driver reads it, believes she has understood, and under-reacts.

• DriveAction.considerStopping said 休憩 — the FATIGUE word. In Japanese driving usage 休憩 is the coffee/rest-stop word (休憩所, 2時間ごとに休憩); it does NOT carry "get off this road, conditions are dangerous". It also contradicted this package's own actionLabel for the same enum value, which correctly says 停車. A driver in a whiteout told 「休憩するという選択肢も あります」 hears an optional comfort break, not a safety escalation — and this is the CEILING rung, the strongest thing this package may say. It now reads 「安全に行える場合は、安全な場所に停車するという選択肢もあります。判断は あなたが行ってください。」 Still an invitation she owns; never a command. (The English was tightened to match the label's weight: "whenever you want one" drained the ceiling rung of urgency.)

• CautionReason.severeAdvisory under-named the top band. That reason covers BOTH AdvisoryLevel.severe and AdvisoryLevel.extreme, and in Japan 警報 (warning) and 特別警報 (emergency warning — the once-in-decades band, a real thing in Akita snow) are distinct terms of art. Rendering both as a bare 警報 lets a driver under-react while a 大雪特別警報 is in force. The band word now admits the top: 「この地域に重大な気象警報（特別警報を含む場合があります） が発表されています。」

  This is a stopgap, and it is named as one: the honest fix is to split CautionReason.extremeAdvisory out of severeAdvisory (the advisor already holds AdvisoryLevel.extreme at the call site). That is a breaking enum change and is deliberately NOT smuggled into a patch release.

Both strings are now pinned by tests that assert their STRENGTH, not merely the absence of a command — so a future edit cannot quietly soften the ceiling rung again.

• feat: DriveAdviceMessages — driver-facing text for the whole advisory vocabulary (DriveAction, CautionReason, Unknown, the compounding note, and the sight-stopping-speed hint), in English and Japanese. Resolve with DriveAdviceMessages.forLanguage('ja'); unsupported languages fall back to English, never to silence.
• Why. Unknown's own dartdoc said these were "an enum, not prose, so the integrator localizes for HER Japanese mother." That was right, and it left every integrator hand-rolling the hardest strings in the catalog, alone, in a language most of them do not read. The wording of a caution at the wheel in a whiteout is not boilerplate to invent under deadline: get one verb wrong and the package that structurally REFUSES to tell her to turn back suddenly tells her to turn back. This release ships the default wording with the package's discipline already baked in.
• The wording discipline is enforced by tests, not by good intentions:
  • continueDriving is the ABSENCE of a raised concern and never an assertion of safety — it says so out loud (「これは路面が安全であることを示すものではありません」).
  • The ceiling considerStopping is an invitation she owns (「選択肢もあります。判断はあなたが行ってください」), never 「停車してください」, and no string in any locale may contain turn-back / abort language.
  • An unknown is never rendered as clear: 「視界の観測データがありません。視界が良い という意味ではなく、データ自体がないという意味です。」
  • 警報 (warning, severe/extreme) and 注意報 (advisory, moderate) are JMA terms of art and are pinned to their correct severities — swapping them would either cry wolf or under-warn a real severe advisory, a defect invisible to an English-reading reviewer.
  • Measured numbers pass through verbatim; the sight-stopping hint states it is a guide and not a guarantee of a safe speed.
• No new i18n mechanism: the hand-rolled locale-table idiom already used by pretrip_decision_advisor's PretripMessages. Pure Dart — the zero-runtime- dependency / 32-bit-ARM contract is untouched.
• test: exhaustive coverage guards (every DriveAction / CautionReason / Unknown has text in every carried locale, so a future enum value cannot ship as a blank caution); a Latin-residue guard that fails if untranslated English leaks into a Japanese safety string; and an end-to-end render of a real compound-failure moment (position lost + whiteout + severe advisory).

• Tests only — the lib/ is byte-identical to the published 0.1.0 (verified by diff against the pub.flutter-io.cn 0.1.0 archive); this release adds regression coverage and changes no production behaviour. (The behaviour bullets that an earlier draft of this entry listed under 0.1.1 in fact shipped in 0.1.0 and are re-attributed there below — they were not new in 0.1.1.)
• Added regression guards on every threshold and the two safety-critical invariants of this D3-worst-case advisor, so a future relaxing tweak fails loudly: (a) exact band-boundary tests pinning each k* edge inclusive/exclusive (visibility 1000/999, 500/499, 200/199 m; freshness 300 s; suspect/degraded radius 150 m; stale-fix 60 s; fast-speed 13.4 m/s); (b) an exhaustive non-deterrence sweep over the full input space including null/NaN/±infinity/ out-of-range values (211,200 combinations) asserting NO input ever produces an action beyond considerStopping — the package is structurally incapable of telling her to turn back; (c) an exhaustive no-under-warn sweep asserting every KNOWN position-concern ≥ 2 paired with a KNOWN visibility-concern ≥ 2 reaches considerStopping AND sets compounding (the D3 stacked-danger core never stays calm); (d) a pin on the deliberate under-confirmation boundary (degraded position + UNKNOWN/STALE visibility holds at heightenedCaution, surfacing both uncertainties, never over-warning a whiteout it cannot confirm); (e) a speed-axis monotonicity guard; and a constants-pinned test that catches any relaxation of the calibration values.

Initial release.

• adviseInDrive(DriveSituation) -> DriveAdvice: one total, deterministic, synchronous pure function that fuses position-trust × visibility (with advisory severity + speed as escalators) into a single honest, immutable in-drive caution record.
• The load-bearing compounding rule: an uncertain position AND low visibility together yields the strongest caution (considerStopping) and sets compounding == true — worse than either alone.
• Three-rung action ladder (continueDriving / heightenedCaution / considerStopping). Structurally no turn-back / abort / do-not-drive rung — the package cannot deter a needed trip.
• Every raised action carries a statable WHY: each escalated action has a non-empty reasons set. The mild single-axis bands each emit one reason — reducedVisibility (~500–1000 m, milder counterpart to lowVisibility) and moderateAdvisory (the moderate level, milder counterpart to severeAdvisory) — and each axis emits exactly one reason (mild XOR hard, never double-counted).
• Sight-stopping hint calibrated to the worst-credible winter surface: it assumes glare/black-ice grip (kWinterDecelMps2 1.0 m/s²) and a winter see-then-react time (kReactionTimeSeconds 2.5 s), plans to stop within HALF the visible distance, and is capped at a winter ceiling (kSightHintCeilingMps, ~48 km/h). It is surfaced ONLY in the low/whiteout band (null in the clear/reduced bands), so it can never emit an "all-clear-ish" figure that overstates the speed at which she could actually stop on ice.
• First-class unknowns: null/NaN/stale visibility (including a reading whose age is null), missing position, and unknown speed surface as explicit Unknown values and distinct CautionReasons, held at a non-zero caution floor and never coerced to "clear".
• suspect position tracks its confidence radius and fix-staleness: a suspect dot at neighbourhood scale, or with no trusted fix for a long blind period, bumps from concern 1 → 2 (matching degraded); a small/fresh suspect fix is unchanged.
• Mirror enums (PositionTrust, AdvisoryLevel) restated locally so the package carries zero runtime dependencies and the integrator maps across the localization_fallback / condition_aggregator / pretrip_source_* seams with one switch per axis.
• Pure Dart, zero runtime dependencies, no Flutter, no IO, no clock, 32-bit-ARM friendly.