fix(ai): align stream TTFT with first effective output - #1301

Summary

Align stream TTFT handling with the first effective streamed output instead of HTTP 200 / header arrival, make explicit null TTFT config wait indefinitely, and update the model settings UI copy to match the new behavior.

Fixes #

Type and Areas

Type:

bug fix / regression fix / UI/UX / test

Areas:

AI adapters, Rust core, web UI, config, i18n

Motivation / Impact

The previous TTFT behavior could keep waiting after the configured timeout in some cases because the timeout boundary was too close to stream setup rather than the first effective streamed output. This change moves TTFT enforcement to the stream handlers, treats text / reasoning / tool-call output as valid first output, removes the reasoning-mode-specific TTFT floor, preserves explicit null as indefinite wait, and updates the settings UI/tooltips so the behavior is clearer.

User-facing impact:

• TTFT now times out based on the first effective streamed output, not HTTP 200 alone.
• Setting TTFT to null correctly means indefinite wait.
• Stream timeout settings copy is clearer, and the UI now notes that changes take effect next turn.

Verification

• pnpm run type-check:web
• pnpm run i18n:audit

Reviewer Notes

• Scope is intentionally end-to-end: runtime timeout semantics, config propagation, fixture coverage, and settings UI copy were updated together to keep behavior and documentation aligned.
• Added fixture/test coverage for delayed first stream output so TTFT no longer treats HTTP 200 as sufficient progress.
• The stream timeout settings UI now uses per-field tooltip hints and a short "changes take effect next turn" description.

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.