feat(crews): compose crews in chat and coordinate directly

[bryantderosier]

Crew creation depended on a /crew composer path, and agent guidance encouraged final handoff files without clearly describing ongoing collaboration. Ask for a crew in ordinary chat now: the current agent proposes saved or custom members, and all members are instructed to exchange findings, blockers, and results directly. Captains request additional expertise through the existing human inbox approval.

Removes the Crew slash-command entry and prompt wrapper. Explicit persona handoff deliverables remain required, but never gate messages. Existing Astra active steering and other providers' safe queued delivery remain unchanged.

Base: crews/10-personas-vocabulary (PR #185); head: crews/11-conversational-collaboration. Entry 11 of stack pingdotgg#153; merge after #185 and before #192.

Validation: focused SQLite communication tests, spawn/prompt contracts, existing delivery integration coverage, scoped server/web typechecks. Native mobile crew approval remains unsupported.

Implemented with GPT-6-Astra (High) in Codex. Reviewed with Claude Fable 5.1 (High) through the Claude harness; the completed stack review found no blockers.

Closes #211.

[Jacksondr5]

Composing a Crew from ordinary chat and telling members to exchange findings directly is the right center of gravity, and this PR is a net deletion on the composer side: the ChatView and ChatComposer deltas remove only the earlier J5 appends, the standing instructions still satisfy every pinned and forbidden phrase in their test, and FORK.md case 8 follows.

One thing to take out. The <j5_crew_launch> parser in crewNotices.logic.ts and CrewLaunchCard in CrewNoticeRenderer.tsx are kept for "stored historical launch notices", and the test inlines the retired fixture to keep them covered. Those notices were only ever written by the /crew command this stack introduced and this PR removes; they exist on dev databases of the unmerged stack and nowhere else. docs/j5/process/working-in-the-repo.md, as written: "Pre-dogfood, no legacy-compatibility code, ever. Accommodations for data or states that cannot exist yet (no users, no legacy drafts) are YAGNI and get deleted." Delete the parser, the card, and the fixture; an old launch turn on a dev database renders as the raw text it is.

Also noted, not blocking: CrewCommunication.test.ts exercises send and proposal storage this PR does not change, so it passes by construction; AGENTS.md asks not to add tests that mirror the implementation.

Reviewed by Claude Fable 5.1 in Claude Code, with an independent pass by GPT-6-Astra in Codex.

[Jacksondr5]

From Jackson's hands-on review, not blocking; fix or defer is your call.

The Captain solicits each member after launch, so every result arrives twice. Right after the launch report, the Captain opened an Exchange with each seat asking for its final joke. Each seat then answered that ask and also sent its result unprompted, as its brief told it to.

The seat brief's <crew_collaboration> block tells members to send findings and results to the Captain. Nothing tells the Captain that. The launch report's footer says "the Captain has each reason" and stops there, and the standing Crew bullet says seats' finishes arrive as messages but not that their results do. One sentence in either place would close it: "Seats are told to message you with their results; do not ask for them, end your turn and read them as they arrive." The report (#148) is the more targeted spot since it is read at exactly that moment.

[bryantderosier]

Addressed the review and duplicate-results comment: removed the retired launch parser/card/fixture and the redundant CrewCommunication test. The launch-report footer now tells the Captain that seats send results themselves, to avoid asking for them again, and to end the turn and read them as they arrive. Focused notice and collaboration tests pass.

[Jacksondr5]

The retired launch parser, card, and fixture are gone, the redundant communication test is gone, and the launch-report footer now tells the Captain that seats send results themselves, not to ask again, and to end the turn. Composing a Crew from ordinary chat with members messaging each other is the right center of gravity, and this PR is a net deletion on the composer side. Approving.

Reviewed by Claude Fable 5.1 in Claude Code, with an independent pass by GPT-6-Astra in Codex.