ADE M8: connect straight to chat + kill the narrative copy

[shuv1337]

Closes #217

The captain's critique: "get rid of this 'resume chatting' button, why would we not connect right to the chat instead of adding this step and the weird narrative in the UI".

1. Direct connect

Opening /fleet/$botId connects. There is no interstitial and no button.

BotChatPage auto-starts once per mount and renders one conversation shell for all three states:

state	conversation region	composer	notice
connecting	pending shell, subtle shimmer	disabled stand-in	—
failed	pending shell, no shimmer	disabled stand-in	compact inline strip: one product-voice sentence + Retry + collapsed Details
ready	real BubbleTimeline / ChatView	live	—

Notes on the two things that made this safe:

• Server-side. startPrimaryChat already takes a per-thread lock and gates on hasSession, so a mount-time start against a bot that is already live is a probe-only pass that adopts the running session. No live-session teardown, no double-start. No server changes in this PR.
• ADE P2: loose bot identity — rename, avatars, groups #197 hook-count gate is preserved. The mount gate machinery stays exactly as it was — ChatView/BubbleTimeline still mount only once chatReady. The pending shell is a pure props-in component that is not an ancestor of either, so it cannot change a hook count. Only the button and the landing page were removed.
• StrictMode. The auto-start effect guards on a useRef keyed by botId, written synchronously before the await. The double-invoked mount effect sees the bot the first pass recorded and issues exactly one start. Keying by botId (not a boolean) lets the same mount reconnect when the rail swaps conversations under it.

The empty conversation is now bot avatar + name + No messages yet — no CTA (the composer below is the CTA), no kernel lecture.

The toolsProbe: "missing" strip is kept (delegation is genuinely down) but reworded and demoted to tone="muted".

2. De-narration

Swept components/captain/** plus the fleet chat/screen copy. Strings stay in the pure logic modules where they already lived; adeCaptainErrorParts is new and splits a captain error into headline + details instead of concatenating them — that concatenation is what put a paragraph of provider setup on screen as primary copy.

Representative before → after:

before	after
Firstmate already has a session open. + Resume chatting	(gone — it connects)
Firstmate is standing by. Nothing is running yet — no session starts until you say so. + Start chatting	(gone — it connects)
This bot runs on the shuvcode kernel. It needs \shuvcode service start` running and an opencode2 provider instance in Settings → Providers (with Binary path pointing at your shuvcode CLI). You can open the chat either way — it will tell you what is missing.`	headline This bot isn't connected — check its provider settings. · Details ▸ the provider/binary remediation
No kernel session is available right now.	This bot isn't connected — check its provider settings.
Start a desktop to watch this bot work. Opening this tab never starts one.	No desktop running (detail empty — the Start button beside it is the action)
You are watching this bot's live desktop.	Desktop running (detail empty)
The desktop is stopped. Its files are kept, so starting it again resumes the data.	Desktop stopped · Files are kept — resuming picks up where it left off.
Waiting for Screenbox to bring the container up.	Starting a desktop (detail empty)
Showing fullscreen. / Paused while off-screen. Scroll back to watch.	Open fullscreen / Desktop running — scroll back to watch
Routines run inside a project. Create one, then this bot can hold routines.	Create a project for this bot.
Send this bot a message once. That opens "Ledger", and routines can be added after.	Send this bot a message to open "Ledger".
Loading this bot's project. Routines appear once it is available.	Loading routines (no detail — nothing to fix)
The rail is showing only the bots waiting on you. Open one to decide it.	(removed — the title states it)
Bots do their work inside a project. Creating one also creates its Second Mate.	(removed — the form is the action)
Code Monkey is mid-turn and this provider cannot steer it. Try again when it finishes.	Code Monkey is still working. Send this when it finishes.
This conversation is still synchronizing. Try again in a moment.	Not connected yet. Try again in a moment.
Sent straight to the fleet. Never shown in this conversation, the contact list, or logs.	Never stored in this conversation.
This bot changed elsewhere while you were editing. Your changes are kept; saving overwrites the other ones.	Changed elsewhere — saving overwrites those changes.
Add one from a template to start building your crew.	Add a bot from a template.
Review work routes to bots whose tag contains "Reviewer".	Tags containing "Reviewer" receive review work.

Scope

M6's attach/observer logic is untouched — only the poster copy in those panels. CaptainShell/ContactRail structure and the workspace sidebar are untouched (M7 owns those); contactRail.logic.ts changes are copy-only.

Verification

• vp test run apps/web/src/components/fleet apps/web/src/components/captain apps/web/src/state/ade.test.ts — 26 files, 414 tests passing
• vp check --fix — 0 errors; no new warnings in changed files
• tsc --noEmit -p apps/web/tsconfig.json — clean
• node scripts/check-brand-identity.ts — passed across 3458 files

New tests: mount-time auto-start incl. a StrictMode-style double-invoke simulation (shouldAutoStartChat), the three-state connect machine + composer gating (resolveBotChatConnectState / isBotChatComposerDisabled), failure-state rendering with the disclosure closed and the technical text kept out of primary copy (BotChatConnect.test.tsx, renderToStaticMarkup — the web suite has no DOM), the error headline/detail split, and copy-tone assertions pinning the banned phrasings out of each rewritten module.