Gate the frame-rate limit on sustained cadence, not the largest gap between two frames

[smb]

An ordinary 60 fps clip could not be opened. The probe gated on the maximum
instantaneous rate between two consecutive frames, and on a millisecond
timestamp grid every 60 fps file alternates 17 and 16 ms gaps because 16.667 ms
cannot be represented. The 16 ms gap reads as 62.5 fps, so one short gap
rejected the file — and told the user to convert it to 60 fps or less, which is
what it already was.

What changed

The gate measures sustained cadence: frames over the duration they span. Two
ticks of the stream time base absorb the floor error quantisation leaves in that
duration — the error bound itself rather than a chosen tolerance.

A fixed frame-rate allowance cannot work here. The error is in the time domain
and scales with 1/frames: 60.6 fps for a four-frame clip, 60.06 at one second,
60.002 at thirty. The first draft of this change used a fixed 1/400 allowance
and would have rejected every quantised 60 fps clip shorter than about 27
seconds — the bug it set out to fix, for the clip lengths a loop tool sees most.

All three comparisons against MAX_SOURCE_FPS move together: the probe, the
declared-rate check in _source_info, and proof verification. A file that opens
now also survives every later decode instead of failing at render time.

peak_rate becomes sustained_rate on both carriers, with its readers in
core/timeline.py, ui/source_presentation.py and the string lookup in
scripts/screenshots.py. The per-pair maximum is removed — nothing read it.

Tests

• test_millisecond_quantized_60fps_source_opens_and_proof_verifies — 1800
  frames, gaps alternating 17/17/16 ms. Fails before the change with
  SOURCE_FPS_UNSUPPORTED, passes after, and exercises proof verification on
  the positive path.
• test_short_millisecond_quantized_60fps_source_opens — four frames, gaps
  16/17/16. The case a fixed tolerance could not cover.
• test_sustained_120fps_source_is_rejected — the limit still holds.
• test_vfr_timestamp_scan_uses_sustained_cadence_even_with_declared_rates
  replaces the test that asserted the old statistic, and still proves the
  timestamp scan runs and governs the outcome.

uv run ruff check ., uv run mypy src, the full suite (1786 passed, 4
skipped) and scripts/check_guardrails.py are green.

This is the fourth consecutive fix: commit touching jobs/source.py; the
maintainer agreed to it explicitly, with an observed trigger and a
previously-failing test.

Closes #68

🤖 Generated with Claude Code