v0.12.4

**Meeting infrastructure you can leave running: this release closes the failure modes behind our worst incidents, and where things still break, it makes Vexa break honestly.**

> 🎉 **New platform — Jitsi joins the lineup.** Vexa now sends bots to **Google Meet, Microsoft Teams, Zoom, and Jitsi**. Jitsi support arrived recently and this release makes it solid: bots correctly handle real members-only lobbies (#597).

If you self-host transcription, reliability *is* the product. A bot that silently drops out of a meeting, a recording that reports "saved" but downloads as zero bytes, a dashboard that shows green while every request fails — these don't just annoy, they cost you the meeting, and you can't get it back. v0.12.4 is focused entirely on making Vexa dependable: bots that stay seated, data that survives crashes, and a system that tells you the truth when something is wrong.

## Bots that stay in the meeting
- **Microsoft Teams:** a bot that got admitted no longer hangs up on itself a fraction of a second later. A background component failing to load used to be fatal; now it's survivable and the bot stays and keeps transcribing (#599).
- **Jitsi:** joining a members-only room now correctly waits in the lobby instead of silently getting stuck in a broken state (#597).
- **Google Meet:** a bot can no longer freeze forever on an unhealthy display (#608), and self-hosters can tune speaker-stream latency with ordinary environment variables (#609).

## Your recordings and transcripts survive
- **Recordings download complete.** A recording the API confirmed as uploaded now downloads byte-for-byte, even if the bot crashed mid-meeting (#611).
- **No more silent transcript loss.** We closed a class of failure where the pipeline quietly froze for hours while health checks stayed green and transcripts were lost — it now reports real health and bounds its memory (#606).

## It fails honestly, not silently
- **Won't boot broken.** A deploy missing a required secret now refuses to start with one clear error naming the missing key, instead of coming up "healthy" and then failing every request (#605).
- **Honest errors under load.** When auth is down, callers get an honest "try again shortly" (503) instead of a misleading "invalid key" (401). Valid keys keep authenticating even when the system is busy — closing the mass sign-in failures some users hit in early July (#601).

## Steady under pressure
- **Database stays responsive** during busy hours: transcript polling no longer holds a connection open while it waits (#602), and a brief Redis blip heals itself in seconds instead of needing a manual restart (#603).
- **No duplicate billing.** Each webhook event now carries a stable ID across retries and replays, so receivers can ignore duplicate deliveries (#604, #519).
- **Guardrails baked in.** New automated checks block any change that would exceed the database's connection limit (#610) or quietly alter its schema (#612).

## Proven on
The **Teams** (#599) and **Jitsi** (#597) fixes were **witnessed live** — a real bot joining a real meeting and staying seated. The **config-safety refusal** (#605) was confirmed on the actual published image. Everything else is **proven by named automated tests**, not assumed — we say what we saw, not a bare "it works." Full per-value witness receipt: `releases/v0.12.4/witness.json`.

**Upgrade:** self-hostable today via the published **`:v0.12.4`** images (stable `:v012` now points here) — you run the exact build we validated.