fix: merge HeZeBang's fork fixes + accessibility/i18n audit items (#128)

[ananmouaz]

What

Two fixes cherry-picked from @HeZeBang's fork, plus the accessibility/i18n items from the 2026-07 audit that were still open.

From the fork (verified locally: flutter analyze clean, 59/59 client tests pass)

• fix(client) — settle dangling tool calls before submit/regenerate/editMessage. A turn interrupted mid agent-loop (stop(), a maxSteps cutoff, or a transcript rehydrated from storage) leaves a ToolCallPart with no matching ToolResultPart; providers reject that history, so every later submit fails. The controller now inserts a synthesized error result immediately after each affected assistant message — in place, not appended, because the Anthropic mapper emits tool results as the following turn and an end-appended result would not repair a mid-history hole. Two tests cover it, including the buried-mid-history case.
• fix(elements) — aiLightHaptic compiles on both upstream Flutter and the OHOS fork. The exhaustive TargetPlatform switch can't cover both enums at once (the fork adds TargetPlatform.ohos), so it's an iOS/Android allowlist now.

Only change made on top of the fork's commits: dart format (they were authored under a different formatter version and would have failed --set-exit-if-changed).

Accessibility & i18n (#128)

• i18n — every string that bypassed AiLocalizations now goes through it: tool-call sections (toolArguments/toolResult/toolError), context meter (contextLabel), source expand/collapse (moreSources(int)/showLess), live-voice statuses, empty-state headline, inline-citation and model-selector semantics labels, attachment remove badge. AiChainOfThought.title, AiContextMeter.label, AiEmptyState.title and AiComposer.hintText are now String? and fall back to the localized value.
• Touch targets — the attachment remove badge and AiBranch's arrows get 44px InkResponse targets (the badge reserves the room via padding: a hit area that merely overhung the Stack would be clipped by the ancestor's bounds check). AiModelSelector's trigger moves from GestureDetector to InkWell for focus traversal and keyboard activation.
• Contrast — AiConfirmation derives the filled label color from the fill's brightness (white on the caution amber was ~2.2:1, AA needs 4.5:1) and uses minHeight + padding instead of a fixed 40px height, so labels don't clip at large text scales.
• RTL — directional alignment for the streaming loader and the "New chat" label; directional padding/border for Markdown blockquotes.
• demo — gallery section header was ~3:1 on white; darkened. Two affected screenshots regenerated.

Versions

• flutter_ai_client 0.3.1
• flutter_ai_elements 0.3.0 — minor, not patch: the four widget fields above widened to String?

Tests

New packages/flutter_ai_elements/test/a11y_i18n_test.dart (10 tests) covers the localization fallbacks and overrides, the 44px labelled remove badge, the caution-tone label color, text-scale growth, and the RTL directional widgets. Full suite: all packages pass (flutter analyze clean repo-wide).

Note on the audit issues

While scoping this I checked every open audit-labelled bug issue against the code. All of these are already fixed on main and can be closed: #109 #110 #111 #112 #113 #114 #115 #116 #117 #118 #119 #120 #121 #122 #123, plus every item in #131 except the "verify claude-opus-4-8 is still served" checkbox (needs a live API call).

Closes #128

🤖 Generated with Claude Code