The capsule geometry cache must reflect only applied layout, otherwise a transient resize or position failure would permanently suppress future retries for the same monitor and translation state. The cache is now written only after the relayout call succeeds, so a failed attempt can be retried on later emits.
Constraint: Keep the fix local to the capsule relayout path.
Rejected: Cache before calling relayout | can strand the capsule on stale geometry after a transient failure.
Confidence: high
Scope-risk: narrow
Directive: Never treat capsule geometry as applied until the window manager has accepted the resize/reposition.
Tested: cargo test --manifest-path src-tauri/Cargo.toml --lib -- --nocapture
Not-tested: Live retry behavior under actual window-manager failures.
The capsule window now tracks its last applied geometry and only resizes/repositions when the monitor or translation state actually changes. That keeps steady-state level updates from churning the native window manager while preserving the current placement behavior when geometry really does change.
Constraint: Keep the fix minimal and avoid introducing a broader window-management abstraction.
Rejected: Leave the unconditional relayout in place | it adds avoidable compositor churn during recording.
Confidence: high
Scope-risk: narrow
Directive: Any future capsule geometry change should update the cached geometry key, not reintroduce unconditional relayout.
Tested: cargo test --manifest-path src-tauri/Cargo.toml --lib -- --nocapture
Not-tested: Live compositor/jitter behavior on Windows/macOS/Linux.
QA positioning had kept its own fixed capsule height constant after the capsule window was resized to match the visible pill. Reusing the shared capsule height source keeps the documented gap intact without widening the change surface.
Constraint: Keep the change minimal and avoid a separate geometry model for QA.
Rejected: Keep the old QA constant | would drift from the new capsule geometry again.
Confidence: high
Scope-risk: narrow
Directive: If capsule geometry changes again, update QA anchoring from the shared capsule source instead of introducing another fixed constant.
Tested: cargo test --manifest-path src-tauri/Cargo.toml --lib -- --nocapture
Not-tested: Live QA window placement on each platform.
The capsule window used a fixed oversized frame even when only the pill was visible, so transparent padding could block clicks on underlying apps. The runtime now resizes and repositions the capsule to the actual visible content, expanding only when the translation badge is shown.
Constraint: Keep the fix minimal and preserve the translation badge when it is actually visible.
Rejected: Add click-through behavior to transparent pixels | too risky and platform-specific for this narrow fix.
Confidence: high
Scope-risk: moderate
Directive: Keep capsule window geometry in sync with rendered content; do not reintroduce a larger transparent hitbox without a matching click-through strategy.
Tested: cargo test --manifest-path src-tauri/Cargo.toml --lib -- --nocapture; npm ci; npm run build; git diff --check
Not-tested: Live mouse hit-testing on Windows/macOS/Linux after the resize change.
The hotword injection copy now tells the model to prefer the user's spelling proactively while still avoiding unrelated mechanical replacements. That keeps the intent of issue #63 aligned with the existing prompt structure and only changes the wording that users see indirectly through polish behavior.
Constraint: Keep the change minimal and avoid changing the surrounding prompt architecture.
Rejected: Broaden the prompt or add new prompt sections | unnecessary for this issue.
Confidence: high
Scope-risk: narrow
Directive: Keep future hotword guidance user-facing and proactive, but still avoid blanket replacement language.
Tested: cargo test --manifest-path src-tauri/Cargo.toml polish::tests -- --nocapture; cargo check --manifest-path src-tauri/Cargo.toml
Not-tested: Live LLM output quality against the updated prompt wording.
Validation could leave a previously fetched model list visible after credentials changed or validation failed, and model selection updated before persistence completed. Clear stale models before validation and only reflect selected models after the credential save succeeds.
Constraint: Address review feedback without changing provider discovery scope.
Confidence: high
Scope-risk: narrow
Tested: npm run build
Tested: git diff --check
baiqing