feat(ai): simplify model selection and add turn usage metrics

[GCWing]

Summary

Model selection and model setup now keep the controls and model context together, and completed chat turns expose their reported usage.

• Expand model modes and reasoning choices in place, preserve keyboard focus and retry context, remember the manual model, and collapse after persistence succeeds. Adapter-reported effort aliases simplify the display while keeping existing preset IDs and request actions intact.
• Organize model settings around acquisition, selection modes, and a searchable model pool with capability/provider filters, default-role badges, per-model tags, and shared action controls. Tags remain annotations and preserve legacy/unknown metadata during edits. Smart/Battery routing and automatic tag synchronization remain explicitly unavailable.
• Show completed-turn token usage, cache-hit ratio, and average output rate. The rate uses whole-turn duration, including tools and waits. Unknown telemetry stays unknown. Shared Tooltips support click pinning and immediate scroll dismissal; historical footers reveal through temporary hover/focus leases without reprojecting the transcript.
• Add the shared IconButton annotation variant and keep Design Lab previews and all three locales aligned with the published components.

Type and Areas

Type: Feature, UI/UX, compatibility, tests.

Areas: Web UI, shared design system, Design Lab, AI adapters, shared AI DTOs, Remote Connect catalog fixture.

Motivation / Impact

Users can configure and find models in one pool, adjust thinking without leaving the selector, and inspect actual turn usage. Existing provider/model/account identity, model defaults, saved reasoning IDs, and request semantics remain compatible.

Verification

PR CI passed all 11 applicable checks; 2 jobs were skipped by the existing impact rules. Full Web UI testing passed 942 files and 10,448 tests.

Passed locally:

• pnpm.cmd run check:web (type checking, Appearance, typography, all-surface color audits, and theme contracts).
• pnpm.cmd run i18n:audit, pnpm.cmd run motion:audit, and pnpm.cmd --dir src/web-ui run icons:check.
• Focused Web UI tests: 24 affected/related test files with 481 passing tests; settings, draft, and picker tests: 30 files with 183 passing tests.
• pnpm.cmd --dir src/web-ui exec vitest run --maxWorkers=50% src/flow_chat/components/FlowChatTypographyRoles.test.ts (6 tests covering shared menu/metric typography and accessible completion details).
• Focused shared UI Node tests for IconButton, Icon, Tooltip, registry, and turn metrics.
• pnpm.cmd --dir design-system/apps/design-lab run build and focused preview/i18n contracts (49 passing tests).
• cargo test --locked -p openbitfun-ai-adapters --lib models_dev::tests (32 tests).
• cargo test --locked -p openbitfun-core-types --lib remote_model_routing_compatibility (2 legacy/current payload tests).
• cargo test --locked -p openbitfun-core --no-default-features --features ai-adapter-runtime --lib infrastructure::ai::reasoning_catalog (11 tests).
• cargo test --locked -p openbitfun-services-integrations --no-default-features --features remote-connect --test remote_connect_contracts remote_connect_model_catalog_builder_preserves_config_shape (1 test).
• Staged/base diff checks, Git object-size policy, and merge-tree conflict check.

Reviewer Notes

Windows desktop library tests hit the existing upstream Tauri test-loader limitation before execution. The workflow accepted its exact documented loader signature; Windows compilation, CLI, and the other platform contracts passed. Linux and macOS desktop library tests passed.

AI-assisted; focused automated testing completed. No browser automation, mock visual acceptance, native/device visual validation, or live-provider inference was performed.

Remote workspace, Remote control, Peer Device Mode, and Detached Dispatch were not exercised live. Target-owned/external selector behavior, optional-field wire compatibility, and the Remote Connect catalog shape have automated coverage; these checks do not establish live remote behavior.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.