feat(ui): select model effort in compare and review dialogs

[astyfx]

Summary

Compare and Review let you choose a model but never a reasoning effort, so both surfaces always ran at the provider's configured tier. This adds a shared effort axis to the model picker, brings Compare onto the same searchable selector Review uses, reworks the Review focus list, and fixes the Compare textarea background.

Changes

Features

• providers: new src/lib/providers/model-effort.ts owns supported effort values per provider/model, settings-derived defaults, cross-model clamping, and the claudeEffort / codexReasoningEffort runtime-override mapping
• ui: ModelSelector gains an opt-in effort axis — the trigger reads model · effort and the search dialog gets an effort row
• review: local change review sends the selected effort as a runtime override for that turn only
• compare: each candidate and the independent judge pick model + effort through the searchable ModelSelector instead of the provider/model dropdown pair
• compare: candidate effort flows into the task runtime overrides, judge effort into the judge runtime options, and both are shown in the compare result panel
• review: focus list expands from 5 to 7 (adds UI & accessibility and Error handling) and each item now carries a one-line description explaining what it instructs the reviewer to do

Fixes

• compare: drop the hardcoded bg-surface on the shared-brief and review-contract textareas so they match the Review dialog in every theme

Refactors / Chores

• review: focus labels, descriptions, and prompt instructions now come from one module (src/lib/local-change-review.ts) so they cannot drift
• store: launchCompareRunVariants takes a runtime-override patch instead of a bare model string, keeping the provider branch out of app.store.ts

Breaking Changes

Test Evidence

bun run typecheck                  # clean
bun run check:max-lines-ratchet    # passed (4 files)
bun run check:switch-exhaustiveness# clean
bun run check:reliability-gates    # passed (6 gates)

bun test tests/model-effort.test.ts tests/model-selector-effort.test.tsx \
  tests/compare-runs.test.ts tests/compare-run-store.test.ts \
  tests/local-change-review.test.ts tests/custom-theme.test.ts
# 66 pass, 0 fail

New coverage: tests/model-effort.test.ts (clamping, provider override mapping), tests/model-selector-effort.test.tsx (trigger renders the effort suffix only when opted in), plus compare/review cases for effort persistence, judge runtime options, and the expanded focus prompt.

A full bun test run also surfaces 11 pre-existing parallel flakes (provider binary discovery, PR-diff timeouts) that pass when their files run individually — unrelated to this change.

Notes

• Effort defaults follow the existing per-model runtime preference, so behavior is unchanged until a user picks a different tier.
• Unsupported carry-over values are repaired rather than sent: a Codex tier steps down to the nearest supported one (Luna ultra → max), and a value with no ladder on the target provider is dropped so the provider's configured effort applies.
• Theme guardrails: no new semantic tokens; the effort row and focus cards reuse existing primary / muted / border tokens, and the textarea fix removes a hardcoded surface instead of adding one.
• Visual verification of the two dialogs was not run in-app; validation is typecheck, repo gates, and the tests above.

[heath-s]

Approved