fix(gemini): stabilize ask reply state handling

[Astro-Han]

Description

Stabilize gemini ask by replacing the old reply detection path with a conservative reply-state model.

This change tightens three parts of the Gemini adapter:

• confirms message submission before reading a reply
• assigns reply ownership from the submitted user turn instead of stale indices or assistant counts
• keeps transcript parsing as a low-trust fallback instead of the primary path

Related issue:
#390

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 🌐 New site adapter
• 📝 Documentation
• ♻️ Refactor
• 🔧 CI / build / tooling

Checklist

• I ran the checks relevant to this PR
• I updated tests or docs if needed
• I included output or screenshots when useful

Documentation (if adding/modifying an adapter)

• Added doc page under docs/adapters/ (if new adapter)
• Updated docs/adapters/index.md table (if new adapter)
• Updated sidebar in docs/.vitepress/config.mts (if new adapter)
• Updated README.md / README.zh-CN.md when command discoverability changed
• Used positional args for the command's primary subject unless a named flag is clearly better
• Normalized expected adapter failures to CliError subclasses instead of raw Error

Screenshots / Output

Checks:

• npm test
• npm run typecheck
• npm run test:adapter -- src/clis/gemini/utils.test.ts src/clis/gemini/reply-state.test.ts src/clis/gemini/ask.test.ts

Manual verification:

• npm run dev -- gemini ask "请只回复：RV1" --timeout 20 -f plain -> 💬 RV1
• npm run dev -- gemini ask "请只回复：RV1" --timeout 20 -f plain -> 💬 RV1
• npm run dev -- gemini ask "请只回复：RV3" --timeout 20 --new true -f plain -> 💬 RV3

[Astro-Han]

A quick follow-up on what this revision improves.

This version does not keep layering more heuristics onto the old path. Instead, it tightens gemini ask around a more conservative submission and reply-state model, with three main changes:

1. Confirm submission before waiting for a reply
   After sending the prompt, the command no longer treats the latest visible text as the answer immediately. It first confirms that Gemini has actually accepted the message.

2. Assign reply ownership from the current-round user turn
   Instead of relying on stale indices or assistant counts to guess which round a reply belongs to, it captures the current-round user anchor at submission time and only accepts assistant content that belongs after that anchor.

3. Demote transcript parsing to a true fallback path
   Transcript parsing is now a low-trust fallback instead of the primary source of truth, which greatly reduces cases where welcome text, page chrome, or partial streamed text gets mistaken for the final answer.

This revision also adds state-driven tests and command-level verification for the main edge cases, especially:

• running the same prompt twice in a row
• the --new true path
• cases where older history is loaded later and must not be mistaken for the current-round reply

In one sentence: this revision makes gemini ask much more reliable at distinguishing between “the message was actually submitted” and “the visible reply really belongs to this round.”

[jackwener]

Review

Well-designed reply-state model that properly separates submission confirmation from response detection. The two-phase approach (waitForGeminiSubmission → waitForGeminiResponse) with user anchor turn ownership is a significant reliability improvement over the old transcript-diff approach.

Fixed: Node-side throw new Error(...) → throw new CommandExecutionError(...) for composer failures (2 instances), matching the PR's stated goal of normalizing to CliError subclasses. Also cleaned up extra blank lines.

Minor note (non-blocking): createPageMock() is duplicated across 3 test files — could be extracted to a shared helper later.

35 tests pass, typecheck clean. LGTM.