feat: add plain-text session suggestions across chat surfaces

[bobleer]

Summary

After a successful interactive turn, offer one optional next-input suggestion using the turn's model and request context. The model returns plain text; accepting a suggestion fills and focuses the composer without submitting it.

• Try the lowest supported reasoning setting, then retry once with the original session client only if preparing or calling the first client fails. Both attempts share an 80-second deadline and cancellation; empty/null output, tool calls and filtered responses remain silent. No JSON Schema, format negotiation, truncation or rewritten output.
• Add the session setting and localized desktop keyboard hint, plus mobile web, Android, iOS and HarmonyOS composer actions.
• Keep suggestions tied to the latest completed turn and device surface. New turns, history rewrites, host restart and disconnect gates prevent stale hints from appearing or crossing sessions/devices.

Type and Areas

Feature and regression protection. Agent runtime/assembly, config/event contracts, remote session streams, desktop Web UI, mobile web and native mobile.

Motivation / Impact

Users can continue a conversation from a grounded follow-up without retyping it. Suggestions are enabled by default and configurable in session settings. The runtime reuses the completed turn's request prefix, tools and cache routing; it never executes suggested tool calls. CLI, ACP, bots, detached jobs and subagents are excluded from generation.

Legacy config defaults and explicit opt-out round trips are covered. Older hosts omit the optional setting/event and remain usable. This adds no RPC command or required protocol capability.

Verification

Passed locally:

• Focused Rust suggestion tests: 9 core tests and 7 portable runtime tests, including real local HTTP/SSE fixtures for plain-text output, reasoning fallback, request identity, cancellation and the shared timeout.
• Six focused Web UI test files: 92 tests covering settings, event delivery, stale-turn retirement, device isolation and composer key policy; Web UI and mobile web TypeScript checks.
• pnpm --dir src/mobile-web run test:host-stream: 22 tests; test:prompt-suggestion-browser: real browser/production composer at 390px and 1200px, exact text, focus, no submission, offline/sending gates and new-turn retirement.
• Focused KMP session/replica tests; iOS simulator core compilation. Android suggestion UI test on compact and wide emulator layouts; iOS suggestion UI test on iPhone and iPad simulators.
• HarmonyOS session-record/composer/host-stream tests: 52 tests, and debug HAP build.
• pnpm run check:core-boundaries, i18n:audit, theme:color-audit:all, capabilities:check, mobile:ui:check, harmony:architecture, formatting and diff whitespace checks.

Existing local failures outside this diff: check:web stops at 25 typography audit violations in design-lab/MobileAvatar/MobileTag; mobile:architecture reports the existing shared test platform-tree reference and three public constructor defaults. The broader KMP session suite has an existing attachment retry failure reproduced with the unchanged baseline. These files/baselines were not relaxed for this feature.

Remote coverage uses simulated host streams and provider fixtures. No live SSH workspace, Relay, Peer Device or Detached Dispatch topology was exercised; HarmonyOS device/foldable visual verification was unavailable. The opt-in live-provider desktop E2E is included but was not used as evidence for this review.

Reviewer Notes

AI-assisted implementation and review. Four feature commits separate runtime, Web surfaces, Android/iOS and HarmonyOS; a merge commit synchronizes current upstream main. Generated capability and mobile design artifacts are included with their owner changes. GitHub CI passed on ed1e8d1ff74c6a15f0fe5b9c43cbdaa43cd568b9: CI run, including all three Rust/CLI platforms and Frontend Build. The Linux full core suite passed 2,590 tests and its Desktop suite passed 431 tests; all nine suggestion generation tests ran successfully.

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.