feat(money-bots): pick a venue + stop the app freezing on inline LLM calls

[rayketcham]

Pick a venue

The board only ever worked whatever the probe surfaced. Now you can point a cycle at one venue — the same move Missions made for the general boards.

• Only US-eligible venues are offered (Kalshi, Polymarket US, Alpaca, Interactive Brokers, Tradier, Coinbase, Kraken). Offering the rest would be offering a cycle the gate is going to refuse.
• A choice always produces work. Tradier and Polymarket US have no SDK repositories, so the probe never surfaces a candidate for them — without a fallback, picking one would answer "no venue program surfaced" and read as a broken picker. program_for_venue grounds those in the venue's own documentation instead.
• Default stays "Any — follow the signal".

The event loop, properly this time

The earlier sweep caught direct backend.call(...) inside async def. It missed the worse case, one frame down: semantic_dedup_check is sync and shells out to the CLI, called inline from async should_accept — so every borderline save froze the entire app for the length of an LLM call. Same shape in /api/launchpad, /api/recruiter, the builder wizard, and the super-idea synthesizer.

The detector is now transitive: sync functions that call the backend, then anything sync calling those, failing on any async function invoking one inline. That's how it caught estimate_build, which never touches a backend itself — it calls _llm_estimate, which does.

Two false positives are documented rather than silenced:

• a factory that returns a blocking callable doesn't block;
• a sync callback defined inside an async function doesn't run on the loop — whoever invokes it decides, and capture_outcome_signals now awaits it off-thread.

Blocking feed fetches are threaded too, and GitHub calls now carry a token when present: unauthenticated allows 60 requests/hour and one probe makes 26, so a few cycles exhausted the quota and every source began returning 403 — which the probes correctly degraded to "nothing found", making the board look dead. The token is only ever sent to GitHub hosts.

The tail narrates the slow part now

Generation was announced only after it returned, so the tail sat frozen on category for the entire minutes-long model call — exactly the stretch being watched. It now announces on the way in, and the test asserts two generate events rather than "a generate stage exists", which is what let the gap through.

The tail also keeps itself current: 15s idle heartbeat, 3s while live, so an open tab picks up runs it didn't start.

Verification

• 2497 passing, 30 skipped; ruff clean
• deployed and verified live: picker renders the 7 eligible venues, board responds in ~14ms

Operational note: --reload watches a bind-mounted path where inotify doesn't fire, so Python changes never hot-load — templates and static files look live because they're read per request. The worker has to be restarted explicitly after Python edits, which is why several fixes in this branch appeared not to work when they were already on disk.

Generated with Claude Code