v0.21.0 — screen recordings are summarized from the screen

No migration. Requires millet-pipeline >= 0.20.0.

A narrated screen recording's iteration plan used to be written by a model
that had never seen the screen. Cue frames are sampled from transcript
timestamps, so they cannot exist while millet transcribe is running —
which is precisely when the summary was produced:

transcribe()      <- summary happened here, no frames yet
extract_frames()  <- frames created here, too late

For these sessions vezir now suppresses the summary during transcription,
extracts frames, and summarizes once afterwards. One summarization, not
two, and the model can see what it is describing.

On a real 31-frame session this surfaced five defects that are not
derivable from a transcript at any model quality — including a destination
field accepting a mainnet address while the wallet was on regtest
(a funds-loss bug), a button bar overlapping a settings row, and three fee
tiers all rendering an identical rate.

Changed

• Deferred summary for video + iteration-plan sessions.
  meet_runner.build_transcribe_args(defer_summary=True) passes
  --no-summarize and omits the preset/template, then
  _run_deferred_summary() reuses millet label --apply-json with an
  empty label map — which millet documents as "re-run just the summary+PDF
  step", the same primitive the retry-summary path already uses. No new
  millet subcommand, no second git sync.
  Every other session keeps the single-pass flow and pays nothing.
• The two summary-failure guards are skipped while deferring, so a
  deliberately summary-less transcription isn't reported as
  "preset requested but no summary was generated".
• A deferred summary that fails is a summary failure, not a job
  failure: the transcript is already on disk, summary_error is set, and
  retry works — the same contract as before the deferral.

Fixed

• millet-pipeline floor was >=0.17.0, two releases stale. vezir
  0.20.0's notes claimed >= 0.19.0 but the dependency was never raised,
  so a fresh pip install "vezir[server]==0.20.0" could legally resolve
  millet 0.17.0 — whose confidential preset points at the retired
  glm-5-2, i.e. an install that 503s on every summary. Now >=0.20.0
  (also required for frames). Existing deployments were unaffected:
  resolvers picked the newest millet anyway.

Tests

1149 → 1160 (11 new in tests/test_deferred_summary.py: which sessions
defer, argument construction, summary-only invocation shape, failure
handling, and an end-to-end ordering assertion that frames exist before
the summary runs).