feat: gemini-acp text streaming compat

[kevin-david]

Why

Gemini-acp's text stream has a few quirks that the adapter wasn't handling. Coding sessions on gemini behaved noticeably worse than on Claude Code: chain-of-thought dumped into the response, long messages truncated to 1900 chars, and tables rendered as raw markdown pipes. This PR closes the gap.

What changes

1. Inline-marker chain-of-thought stripping. Gemini emits reasoning as plain text with a literal [Thought: true] marker at the thought/response boundary. At medium and below we retroactively trim the draft to only the post-marker content so the user sees the response without the thinking dump. High keeps both visible. Adds a MessageDraft.replaceBuffer helper for the in-place rewrite.

2. Finalize text draft on every turn end. Agents that don't emit usage / session_end at turn-end (gemini) left the streaming text draft stuck at its 1900-char mid-stream truncation. Register a plugin middleware on the turn:end hook that calls a new public DiscordAdapter.finalizeSessionDraft so the buffer gets split into chunks regardless of the agent's terminal-event behavior. Also finalize on incoming user message (messageCreate) so the prior turn's draft seals before the next turn's text appends to it.

3. Discord-specific rendering instruction injected per session. First-prompt-only middleware on agent:beforePrompt prepends a <system_instruction> block telling the agent: no markdown tables, use ASCII tables with +---+ borders inside triple-backtick fences, tables ≤90 chars wide, apply silently. Discord-specific (only fires for sourceAdapterId === 'discord'). Adds the middleware:register permission to the plugin. (Supersedes #10)

Stacking note

Works best in combination with #11 (content-loss split fix) — without that, finalize would still drop content into the void on long responses. Order of merge: #11 first, then this.

Test plan

• Long table responses now arrive in full across multiple Discord messages
• Gemini chain-of-thought no longer dumps into the response at medium
• First-turn instruction appears to be honored (no markdown tables, ASCII style with borders)
• Reviewer: confirm assistant session (PR feat(assistant): instruct agent to fence tables and fixed-width content for Discord #9 already shipped) and regular sessions both render tables correctly
• Reviewer: confirm no regression for Claude Code (no [Thought:] marker → retro-rewrite is a no-op)

🤖 Generated with Claude Code

[kevin-david]

Check that this doesn't break with back to back messages

[kevin-david]

back to back messages are fine - anything else to be worried about here?

[kevin-david]

I actually think it's even shorter than this, on my Pixel Fold I can only get 51 characters on folded mode and 71 on unfolded mode. The desktop in full thread view mode is like 91. Thoughts?

[0xmrpeter]

Clean implementation of both features — the replaceBuffer approach for real-time thought stripping is elegant, and hooking into turn:end for draft finalization is exactly the right place for this. The middleware pattern is consistent with how the rest of the adapter works.

A few minor things worth addressing before merge:

• Deprecated getter — finalizeSessionDraft uses session?.threadId which is the deprecated single-thread getter. Should be session?.threadIds.get('discord') to stay consistent with the multi-adapter model.
• Wrong field for adapter check — session?.channelId === 'discord' in the turn:end handler uses the creation-time channelId field. session.attachedAdapters.includes('discord') would be more accurate, especially in multi-adapter scenarios.
• Empty edit guard — replaceBuffer('') when post-marker content is empty will fire an empty Discord edit call. A if (postMarker) guard would skip the unnecessary API call.
• Silent catch — the catch in the turn:end middleware swallows errors completely. Even a log.warn would help with debugging if finalizeSessionDraft fails.

This stacks on #11 — worth merging that one first. No blocking issues here.