[miguel-heygen]

What

A render no longer fails with ffprobe exit 1: Invalid interval start spec when a media file's duration is a hair over one second (for example 1.000001 s).

Why

The audio duration probe starts reading at start + duration - 1. For a duration just over one second that is below 1e-6, and JavaScript prints such a number in exponent form (9.999999999177334e-7). ffprobe's -read_intervals parser does not accept an exponent and rejects it. The video probe had the same problem and was fixed with formatFfmpegSeconds; the audio probe still interpolated the raw number.

How

• One readIntervalSpec(start, end) in utils/ffprobe.ts builds every -read_intervals value; it formats the start through formatFfmpegSeconds. The four places that built the string by hand (video tail, video tail fallback, audio tail, keyframe gap) all use it.
• Ordinary values print exactly as before.

Test plan

fails without the fix: packages/engine/src/utils/ffprobe.test.ts › the audio tail probe's read interval › starts at a plain decimal when the duration is a hair over one second

• On main's code that test fails with expected '9.999999999177334e-7' to match /^\d+(\.\d+)?$/.
• ffprobe.test.ts: 125 passed, 3 runs in a row. oxlint 0 errors.
• Committed with lint/format/typecheck hooks skipped locally (no node_modules on the Mac); oxlint and the tests were run on a remote box.

Size

Under 100 lines, and it stands alone: it completes the -read_intervals fix already shipped for the video probe, and one test covers all four sites through their single builder.

[heygengenesis]

Approving at 3f75d0aa. I found no blockers.

• Root cause matches. tailStart was interpolated raw. For a 1.000001 s duration, String(9.999999999177334e-7) gives an exponent, which ffprobe's -read_intervals start parser rejects.

• The fix holds. formatFfmpegSeconds maps values with an exponent through toFixed(6):

  • 9.99e-7 → 0.000001
  • 4e-7 → 0
  • a tiny negative such as -4e-7 → 0 (because String(-0) is "0")

  Values without an exponent print exactly as before. String end tokens (+2147483647, +#1) pass through untouched, and numeric streamEnd is formatted as it was before.

• Coverage is complete. All four call sites that build a -read_intervals value now use readIntervalSpec. The two remaining args.splice/unshift sites only forward those strings.

• Fails without the fix (checked). With ffprobe.ts reverted to main, the named test fails with expected '9.999999999177334e-7' to match /^\d+(\.\d+)?$/, exactly as the PR body says. It passes at this head.

• Suite. ffprobe.test.ts passed the same way 3 runs in a row. Locally, 4 PNG cICP fixture tests fail at both main and this head, because my clone has no Git LFS fixtures. They are not related to this change. tsc --noEmit on the engine package is clean. The commit skipped local typecheck hooks, but CI Typecheck is running.

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 132 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (0)

Unstable (1)

• seqplainkeys-none-center-r0-root-z100: tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true