feat(chat): open the chat in an editor tab - #70

Answers "can the LevelCode AI panel move to the centre?" — by dragging, no, and
that is not our choice: `ViewContainerLocation` is Sidebar | Panel | AuxiliaryBar
and nothing else, so a VIEW can go left or to the bottom but never into the
editor grid. Explorer and Terminal cannot either. Only EDITORS live in the middle.

So the centre needs a WebviewPanel rather than a WebviewView: a real tab that
splits, moves between groups, and drags to another window. `AI: Open Chat in
Editor`.

It is a MOVE, not a mirror. Two live surfaces would mean fanning out every
post(), making every handler idempotent, and shipping a UI that can disagree with
itself; one live surface keeps a single source of truth and is what "open in
editor" means to a user. The sidebar hands its slot to a small card offering
"Bring it back".

Three things carried the design:

**The listener survives an html swap.** onDidReceiveMessage binds to the WEBVIEW,
not the document, and the sidebar's html is swapped between the chat and the
hand-off card. So wire() (register the handler, once) is split from makeLive()
(become the live surface, load the chat). Wiring on every swap would stack
handlers and double-send every message.

**The transcript is DOM state.** A hand-over would land you in an empty chat
holding a conversation the model still remembers. Every transition arms a replay
of the live session's turns, consumed on `ready` — the earliest a fresh webview
can receive anything — and cleared in the same step so it cannot repeat on a
later reload.

**One restore path.** "Bring it back" disposes the panel rather than restoring
the sidebar itself, so closing the tab and clicking the button run identical
code. A sidebar that was never resolved is revealed instead of written to.

The replay reuses the sessionResumed renderer rather than a second one, now
labelled by the caller: a move says "Moved to the editor", not "Resumed" — which
would claim the session had been reloaded from disk when nothing of the sort
happened. A real resume passes no tag and still reads "Resumed".

Tests: 12 in test/chatSurface.test.js, read out of the shipped extension.js the
way mcpManage does, because every failure mode here is state and invisible in a
diff. Verified non-vacuous — each bypass is a bug someone could plausibly write:

  wire inside makeLive (duplicate handlers)   2/12
  drop the already-open guard (two panels)    1/12
  replay not cleared (stacking transcripts)   5/12
  reattach restores directly (two paths)      7/12
  hard-code "Resumed" (a move mislabelled)   10/12
  drop the empty-transcript guard             6/12

33 suites green.

The command shipped palette-only, which is the wrong affordance for it: nobody
searches the Command Palette for a capability they do not know exists, and "the
chat can move to the centre" is exactly that kind of capability. The icon was
already declared and doing nothing, because no menu contributed it.

Now a fourth navigation action on the chat view title, beside Add context, New
Chat and Set API key. `navigation@4` keeps it inline and lets VS Code overflow it
into the … menu when the sidebar is too narrow for four — no crowding, and the
sidebar is narrow by definition until you use this.

Scoped `view == levelcodeAi.chat`; without the when-clause a view/title action
appears on every view title in the window.

One more test, verified non-vacuous three ways — remove the menu entry, drop the
when-scope, remove the icon: 12/13 each. 33 suites green.

Review, and correct: detachedHtml() served a webview document with an inline
<script> and no Content-Security-Policy, while getHtml() and getSessionsHtml()
both inject CSP + nonce. Small is not exempt — the card enables scripts and
carries one, so it was the single document in the extension whose script surface
was undescribed. Worse than untidy: a later tightening elsewhere would have
silently stopped its button from working, and "Bring it back" is the only way out
of the detached state short of closing the tab.

Rather than a third copy of the policy, the construction moves into webviewCsp()
and all three documents call it. It had already been duplicated twice; a third
would have made "tighten the CSP" a three-file change with one easy to miss.

Verified by rendering the card and reading it back:

  CSP          default-src 'none'; style-src 'unsafe-inline'; script-src 'nonce-…'
  script nonce matches the policy
  no remote origins allowed

Two guards, verified non-vacuous four ways — drop the meta (12/15), unnonce the
script (12/15), hand-roll a third copy of the policy (13/15), bypass the helper
in getHtml (13/15). One of them asserts the policy string exists exactly ONCE in
the file, so the next document cannot quietly hand-roll its own.

33 suites green.