R41 Refactor — Pixel-Diff Verification

Artifact: verify_r41_pixel_diff.py Lane: Refactor (behaviour-preserving DRY extraction) Verdict: PASS — refactored r41 card is pixel-for-pixel identical to the original.

What was being verified

/root/.hermes/ops-pulse-card-r41.py is an original, inline-PIL ops-pulse card (tick 2026-10-15 20:59Z, "Sentinel onboarded; model claim verified"). The refactor delegates all PIL boilerplate to the shared ops_pulse_card_lib.render_card() so the card reduces to a single CardParams(...) + render_card(params) call. The library doc guarantees "Pixel-identical output."

The verification renders both the original and the refactored card (both write to the shared /root/.hermes/ops-pulse.png), captures each to a separate file, and compares every pixel with numpy. Exit 0 on identity, exit 1 on any mismatch or run failure.

The bug caught (data lost during DRY extraction)

The first extraction into _extracted_params.json (and the hand-written refactored card that followed it) silently dropped two pieces of the original:

The verdict subtext line — original renders d.text((46, 134), "new 8th wizard "the Long View" added to STANDING-AGREEMENT + WIZARDS-TABLE", font=font(17), fill=MUTED). The refactored card had verdict_subtext='' (empty) and verdict_subtext_font_size=16 (should be 17).
The first row — original's first row is ("Wizard that moved", "Sentinel — "the Long View" (new party, onboarded)", BLUE). The refactored card listed only 5 rows (it started at "Claim verdict"), omitting this BLUE row entirely.

Measured evidence (baseline = original render)

Before the fix (refactored as shipped in the profile):

FAIL total pixel-diff=8295407
diff bbox x[28..972] y[137..508]
diff pixels=72112 of 700000

The diff band y[137..508] spans exactly the verdict subtext slot (y~134) plus the entire row block — because dropping the subtext shifted nothing visually subtext-wise, but dropping the first row shifted every subsequent row UP by y_step=54, cascading a full-column mismatch.

The fix

In refactored_cards/ops-pulse-card-r41_refactored.py:

verdict_subtext set to the original's subtext string;
verdict_subtext_font_size corrected 16 -> 17 (matches original font(17));
Row('Wizard that moved', 'Sentinel — "the Long View" (new party, onboarded)', BLUE) re-added as the first row.

No geometry constants changed — first_row_y=190, row_height=48, ellipse_y0=19, ellipse_y1=31, y_step=54, value_x=330, etc. all already matched the originals (they were correct in the extraction defaults).

After the fix

PASS r41 pixel-identical (1000x700) — original vs refactored_card via ops_pulse_card_lib
exit=0

Total pixel-diff = 0 across the full 1000x700 canvas (700,000 RGB pixels).

Reproducer / CHECK

python3 /root/.hermes/wizard-outputs/forge/verify_r41_pixel_diff.py

2.9 KiB Raw Blame History