[eugene-mux]

Description

Inspect compressed AVC/HEVC video samples for Standard Input planning. Complete scans supply cadence, bitrate and GOP evidence; incomplete scans retain metadata and leave sample facts unknown.

Changes

• Use MediaExtractor to measure frame rate, presentation timestamps, video-average and per-GOP bitrate/bytes, and keyframe intervals separately from metadata inspection.
• Inspect sync-sample NALs for IDR/open-GOP evidence. Ignore non-VCL signaling while keeping malformed headers and unsupported syntax unknown.
• Classify video edit lists by track ID without applying their timeline. ISO parsing is limited to edit/timeline work; Part 1 reads platform metadata directly. Malformed lists preserve track metadata; zero-duration edits remain unknown and Matroska reports no ISO edits.
• Document valid B-frame timestamp reordering and report capped legacy sample reads as LimitExceeded.
• Bound scans to 4 MiB per sample, 250,000 samples, 1 GiB read and 5 seconds of cooperative elapsed time. Limited or interrupted scans discard all partial sample facts; without another known violation, the planner can retain the original file.
• Leave effective duration and A/V offset unknown for the future changes. Trimmed preroll retains unknown bitrate and final-GOP interval when the endpoint is unproven.

Validation

• 124 tests pass in each debug/release variant on JDK 17.
• SDK and sample release builds, lint-vital, whitespace checks and CI pass.
• All 23 external API 36 emulator cases match expected outcomes: 19 complete scans, three unsupported inputs and one malformed-container rejection.
• The legacy read branch matches on API 36; actual API 23 runtime validation remains pending.
• Large 1–3 GB 4K measurements on physical mid-range hardware remain pending. The current caps can leave large-file bitrate/GOP violations unknown; emulator timings do not establish device performance.

---

Note

Medium Risk
New binary parsing and MediaExtractor loops affect upload standardization evidence; behavior is defensive (unknown/partial discard) but incorrect GOP or edit classification could change planner decisions.

Overview
Adds compressed-sample inspection for Android standard-input planning: MediaSampleInspector walks single-track AVC/HEVC via MediaExtractor, derives cadence, timestamps, average/per-GOP bitrate and keyframe spacing, and inspects sync-sample NALs for IDR vs open-GOP structure. Scans are capped (sample size, count, bytes read, elapsed time); partial, encrypted, or interrupted runs return status without publishing sample-derived facts.

ISO edit-list classification is new in IsoContainerMetadataReader: it reads moov/trak identity and classifies elst as none/simple/complex by track ID for video edits only—no timeline application. Matroska is treated as no ISO edits; malformed edit lists leave editList unknown while preserving other track metadata.

TimestampFacts renames isMonotonic to presentationOrderMatchesDecodeOrder and documents that B-frame reordering is expected, not invalid PTS.

^{Reviewed by Cursor Bugbot for commit b037dcc. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 29f8f40. Configure here.}