# What was inspected; what remains different

All 57 positions of the newly decoded reference clip, retained native PNGs and both
newly rendered candidates were inspected in labeled full-composition sheets.
All 57 native pointer/replay crop pairs, the dialog/focus boundary crops and the
excluded S3687 boundary were also inspected. Full-frame sheets are fitted views;
the pointer/focus crops preserve native pixel scale. These are still inspections,
not perception of continuous playback, and not a blind test.

The new replay versus retained candidate result is exact in bytes, all decoded
RGB24/YUV420P frames and all PTS/durations. No changed styling was needed for
portability. This does not erase differences between the **candidate and source**:

- The source's intermediate pointer has different streak shape and black-body
  softness, visibly around L6–14 and L40–55. The candidate keeps the pre-existing
  49-sample exposure and separate head/tail optics rather than declaring them a
  source-equivalent blur. No numeric source-pointer fidelity score was invented.
- The focused source border is blue-toned; the original candidate's retained
  border is green `#2f8175`. Its visible stroke weight/edge treatment also differs.
  The timing of caret on/off states is corroborated; matching blink timing is not
  border or typography fidelity. No color restyling is performed in this lane.
- Source thumbnails and logos are not candidate assets. The authored repeated
  thumbnail scaffold, original diamond icons, sidebar labels, owner details and
  domain strings differ intentionally; they do not establish a quality win.
- Row-four highlight begins while the arrow remains lower. This is present in
  native source pixels, not an invitation to “repair” highlight with hover logic.
- The native field is focused when the dialog appears. The arrow's subsequent
  leftward traverse ends by the access controls. Do not invent a later field click.

The source PNG crop is pixel-confirmed against three retained full-frame witnesses
and y32 centering is rejected. Its custody remains inherited; there was no fresh
master extraction. The encoded clip/native PNG differences are retained per frame
in `native-proof.json`, rather than silently treating MP4 pixels as lossless native
evidence.

The prepared-plate CPU diagnostic compares all 57 positions for both variants.
It is **0/57 exact RGB frames** for each, with maximum channel error 72 and mean
absolute channel error 0.3249666504251328 (original) / 0.32369234380599454 (adaptation)
on the 0–255 scale. These are real differences, not a passed threshold. The
comparison mixes PNG alpha quantization/CPU compositing with a lossy encoded
reference; individual causes were not separated. All per-frame values remain in
`bridge-diagnostic.json`. This diagnostic neither replaces the exact full-scene
Remotion replay result nor establishes a native Fusion composite result.

The fresh independent critic is separate from this builder. This document neither
supplies an anonymous ranking nor preempts that critic. Historical open findings
in `addon/SOURCE.json` remain historical, not newly passed checks. No source-quality,
moving, blind-win, completed-library or native-Resolve credit is awarded here.