fix(mcp): credit the agent on the board, let it tidy up, and survive a restart

[praxagent]

Three things found by connecting a real MCP client and using it.

The board said a human did it

Cards filed over MCP were credited to the person reading the board — which defeats most of what a shared board is for.

Credentials now carry a display label — "Claude Code" by default, overridable so Codex users don't see the wrong tool — and every mutation passes it as the author, editor or comment actor. An unlabelled key falls back to its own name rather than to "human": a key identifier on a card is ugly; crediting the wrong party is worse.

A comment that silently vanished

comment_on_task sent {comment, author}. The Library API reads {text, actor}. Comments were accepted, reported successful, and arrived empty and unattributed — a silent wrong result, worse than an error because nothing tells you to look.

There was a test asserting the broken shape. It checked what our code sent rather than what the API reads, so it passed for as long as the feature was broken. Rewritten against the contract with a note saying so — that's the more useful lesson than the bug.

delete_task added: an agent that files a card should be able to withdraw it. Requires task.write, respects space scope, refreshes the board, and its description says to only remove cards it created or was asked to — a tool the agent can see is a tool it will try.

A tab that never came back

Reported as a WebSocket 502. The 502 was transient — tailscale serve answers 502 while the upstream restarts, which is every deploy. But the reason a tab stayed dead is ours: backoff was 1000 × 2^attempts with no ceiling and a hard stop at 10 attempts.

Attempt	Wait
7	2 min
9	8.5 min
10	gives up permanently

• Capped at 30s with jitter. Backing off protects the server; waiting minutes to notice it came back protects nobody.
• Never gives up. A server down now is usually back inside a minute, and a tab left open overnight should be alive in the morning. "Refresh the page" is the app admitting it stopped trying.
• Refetch on reconnect after a drop. Nothing replays missed events, so a recovered tab showed pre-gap state — the dangerous kind of stale, because it looks fine. Same class as the no-polling Kanban.

Verification

294 backend + 37 frontend tests green, tsc clean. Prax half is praxagent/prax#186.