rsv-ben: fix fps detection for CFR reference clips

[kevinthelobster]

Problem

-rsv-ben recovery of Sony recording-in-progress (.RSV) files produces, for NTSC-fractional footage (23.976 / 29.97 / 59.94 fps):

• one corrupt video frame per GOP — visible as a hitch roughly every half-second, and
• audio that ends short of the video (~4.5 s over a 75-minute clip).

This looks like an inherent limitation of recovering an unfinalized file, so it's easy to misattribute as unavoidable data loss. It is not — the frame data is intact in the .RSV; it is mis-spliced.

Root cause

repairRsvBen() derives the video frame duration only from Track::times_:

if (!video_track.times_.empty())
    video_duration_per_sample = video_track.times_[0];

But Track::getSampleTimes() collapses a single-entry (constant-rate) stts into constant_duration_ and leaves times_ empty:

if (entries == 1 && nsamples1 > 500) {
    constant_duration_ = stts->readInt(12);   // times_ stays empty
}

Every Sony camera records CFR, so the reference clip's times_ is empty and video_duration_per_sample keeps its 1000 default. fps is then computed as 24000 / 1000 = 24 instead of the true 24000 / 1001 = 23.976.

That feeds the per-GOP audio chunk size:

audio_samples_per_chunk = gop_duration_sec * audio_sample_rate

giving 12 / 24 * 48000 = 24000 samples (96000 bytes) instead of the correct 12 * 48000 * 1001 / 24000 = 24024 samples (96096 bytes) — off by ~96 bytes per GOP. audio_boundary = next_rtmd_start - total_audio_chunk_size is then ~96 bytes too late, so the last video frame of every GOP swallows ~96 bytes of PCM (corrupt frame), and the audio chunk is ~96 bytes short every GOP (cumulative drift).

Empirically confirmed on a 58 GB FX30 XAVC S .RSV: at the 96096-byte boundary the bytes before are high-entropy H.264 and after are low-amplitude 16-bit PCM; at 96000 both sides are PCM.

Fix

• Read constant_duration_ when times_ is empty (the actual fix).
• Compute audio_samples_per_chunk with exact integer rounding so float truncation cannot drop a sample at other rates.

Verification

Patched build, same 58 GB FX30 .RSV + a same-codec CFR reference:

• derived parameters: fps=23.976, GOP duration=0.5005s, audio chunk=96096 bytes (was fps=24 … 96000)
• Duration of avc1 and Duration of twos now identical (were 4.5 s apart)
• continuous decode: 0 per-GOP errors (was ~1 corrupt frame per 12-frame GOP); only the genuinely truncated final ~2 s (point of recording interruption) is damaged.