Card and audio-less-clip anullsrc sources are infinite; concat never
terminates an infinite segment, so the encode hung forever at 30%. atrim
pins each bed to its item's span (card duration / clip trim).
Per the adopted qwen38 plan §2 (+ glm53 amendments): video_processing_jobs
kind 'assemble' (claim widened via 044), EdlPlan mirror validation
(lenient-fail), the filter_complex builder (interleaved concat pairs, fit
cover/contain chains, absolute-timeline PNG overlays with enable windows,
VO adelay/sidechain duck, hard-breaking word wrap for card PNGs), and the
runner's assemble path with pre-spawn trim gates, attempt backoff and 15-min
cap. Sources open by magic (SPENC1 kek / SPENC2 direct key) so direct-to-R2
uploads assemble without the api ever seeing bytes. Tests: builder 18,
edl-plan 12, overlay-png 3, runner fakes extended.
downloadSealed fetched raw bytes (getSealedObject) and handed ciphertext to
ffprobe — every direct upload failed 'Invalid data found'. The lane now uses
getObject, whose magic branch opens SPENC1 and SPENC2 alike.
POST /video/direct/init|complete: the browser seals (SPENC2) and PUTs
straight to object storage; the api HEAD-checks, registers the unprobed
asset and queues a probe job — video bytes never cross the api box.
Worker side: produce_job_claim widened to probe jobs (042), the runner
gains a probe lane (sealed fetch → SPENC1/SPENC2 open by magic → scratch →
ffprobe → metadata write-back → complete 100), worker storage reads SPENC2
frames via DIRECT_EC_PRIVATE_KEY (inherited through compose appenv), and a
ported ffprobe util. Probe rows carrying input_json fail instead of
stranding; the claim guard no longer strands probe rows as invalid.
Audio-less inputs no longer emit [0:a] maps that ffmpeg rejects; the worker
branches to an anullsrc bed configured at the output preset's rate/layout.
Producer tests 42/42.
The store's hand-built claim interpolated the lane predicate into the SQL
text while still passing a bind parameter — node-postgres sent 1 param for
a 0-param statement (08P01) on every poll, so the live produce worker
store-errored every ~2s and never claimed. Migration 040 replaces the
inline claim with produce_job_claim(p_lane, p_stale_after): SECURITY
DEFINER (the 035/oidc_refresh_lookup pattern — the poller is lane-global,
FORCE RLS hides every row without a tenant context), atomic FOR UPDATE
SKIP LOCKED with the queued→transcoding bump as the lease, and stale-
transcoding reclaim so a killed worker's rows return to the queue.
produce_job_workspace(p_job_id) discovers the owning tenant for the
follow-up writes, which stay RLS-bound as sp_app. pg-store rewired; store
errors now surface err.message instead of err.name; pg-store tests added.
38/38 producer tests green.
Deployment (t4g2): media-worker gains font-dejavu (drawtext lower-thirds)
and a dedicated produce-worker compose service running src/produce.ts —
claims video_processing_jobs compose|mix by lane, renders assembled/
overlaid/voiced cuts, registers derived library assets. Deliberately its
own process so a remote-lane box needs only Postgres + S3 + ffmpeg.
Web: compose media attachments UI (t2), Producer panel in the video
drawer (t5), collective creation + branding settings + shell branding
(brandweb), mobile CSS (mcss: touch targets, icon sizing, safe areas).
