# Source, crop and clocks

## Identity and interval

The retained source metadata identifies OpenAI's *Introducing Sites in Codex*,
video ID `VRvC5smyzso`, publication date **June 5, 2026**. This identity/date comes
from the retained metadata, not a fresh fetch. Its official locator and custody
hashes are preserved in `addon/SOURCE.json`. Private media paths are not included.

The study contains display-order source frames **[3630,3687)**: exactly 57.
Local frame `L` corresponds to source frame `3630 + L`, for `0 <= L <= 56`.
No frame 3687, subsequent pullback, new opening hold or closing fade belongs here.

## Measurements versus nominal arithmetic

| Clock | Basis | Values |
| --- | --- | --- |
| Standalone reference clip | Newly probed, all 57 frames | 1840×1016, silent video; time base 1/15360; PTS `512*L`; each duration 512 ticks |
| Both actual replays | Newly rendered and newly probed, all 57 frames | Same dimensions, PTS, durations and time base as clip; no audio |
| Source master | Retained probe ledger, not freshly re-probed | 1920×1080; first included PTS 1858560; last included PTS 1887232; excluded boundary PTS 1887744 |
| Nominal local clock | Arithmetic, not an independent measurement | `L/30` seconds; last frame starts `56/30`; interval length `57/30 = 19/10` seconds |

The measured local exclusive endpoint is **29184/15360 = 1.9 seconds**. The last
included frame starts at **28672/15360**, not 1.9 seconds. Rounded decimal probe
strings alternate between apparent 0.033333 and 0.033334 starts; use integer PTS,
not rounded strings, for equality. The master interval is `[121,122.9)` seconds.

The retained `native-audit.json` clock and packaged `addon/SOURCE.json` clock agree
for all included frames and both neighboring records. The master itself was not
available for a new probe in this builder lane. Unlike an extrapolated endpoint,
the retained ledger includes a decoded S3687 PTS/duration and there is also an
S3687 native PNG. This is **retained measured evidence**, not a newly extracted
master clock. `controls/timeline.csv` distinguishes local PTS from inherited
master PTS and from nominal seconds.

## Pixel-grounded crop

The crop is `(x=40, y=64, width=1840, height=1016)` from a 1920×1080 frame:
right-exclusive 1880, bottom-exclusive 1080. It removes 40 pixels at each side and
64 at the top; it is not vertically centered. The alternative `(40,32)` is wrong.

Available retained source evidence contains 58 native cropped PNGs, S3630–S3687,
and three full 1920×1080 PNG witnesses at S3630, S3655 and S3687. Fresh pixel
comparison of each full witness cropped at `(40,64)` against its native cropped
PNG gives zero channel error. Cropping the same witness at `(40,32)` fails all
three comparisons. `native-proof.json` records hashes and actual error values.
The 57 included native PNGs were inspected; the excluded one was examined only
as a boundary witness. None was newly extracted or invented by this builder.

The reference MP4 also decodes at 1840×1016; it is not an upscaled small preview.
Its RGB pixels differ from the retained PNGs because the clip is an encoded
derivative. Per-frame errors are reported, not treated as a lossless native PNG
substitute. Full composition contact sheets are fitted inspection aids, never
called native pixels. Pointer and field crops were inspected at native scale.

## Observed sequence and its reconstruction

- **L0–5 / S3630–3635:** the fifth row carries the pale hover. Pointer starts near
  `(1209,612)` in the reconstruction; that is an authored fit, not a newly
  measured subpixel source coordinate.
- **L6 / S3636:** hover jumps to row four while the pointer is still near row-five
  height. Native interior samples corroborate the row change. Do not infer row
  highlighting by hit-testing the current arrow location.
- **L15–24:** the target Share pill has a separate hover state. Source-native
  pointer crops show approach and optical smear; the authored rounded-target
  hit-test ledger is in the addon. That ledger is not source UI event telemetry.
- **L25 / S3655:** scrim and sharing dialog appear in one frame. Pointer stays over
  the Share pill, above the dimmed background; it does not teleport into the
  dialog. The field is already focused and its caret visible.
- **L25–39:** caret stays visible. Pointer scale recovers through L35; the pointer
  remains near `(1445,531)` through L37, then traverses left.
- **L40–54:** caret is absent, but the field border remains focused. At L40 the
  reconstruction switches its black arrow optics to separate overlapping head
  and tail branches; shadow and white rim remain present.
- **L55–56:** caret returns. The pointer has crossed into the modal and ends at
  authored `(838,543)`, near the access controls, **not inside the input field**.
  Do not invent a click into the field at this point.

For every native PNG, `native-states.json` records empty row-interior RGB samples,
the background scrim sample and dark-pixel counts at the caret column. They agree
with these row, scrim and caret boundaries. This is a narrowly defined pixel
check, not an exact source pointer-path fit or a general source-quality pass.