Unify poster HTML, PDF, PNG, and Reel pixel fidelity

[LingAo Xiao (ArmandXiao)]

Summary

• unify Fix poster HTML and Reel pixel fidelity #39's durable HTML, PDF/PNG, PPTX, and Reel fidelity pipeline with Fix poster figure shrinkage during HTML backfill #40's metric-compatible Arimo mapping and pre-freeze figure-regression gate
• migrate fresh Arial, legacy Arial/DejaVu, and local-first Arimo posters to packaged-first Arimo while preserving media identity, enforcing the absolute 90% research-figure fill floor, and rejecting new axis overflow above 101%
• persist renderer-derived geometry across warm reloads and brand-new browser contexts, then require decoded-RGBA equality before promoting HTML, PDF, and PNG artifacts
• make logo packing rerender-idempotent for both new width-cached zones and completed legacy zones that only stored the old height cache
• keep Reel canonical poster copies and inline srcdoc identical over HTTP and direct file://, with the wordmark bundled as a runtime asset

This PR supersedes #39 and #40 and is the only one that should be merged.

Root cause

PR #37 made the standalone HTML reload path durable, but Arial was frozen to DejaVu Sans before the source baseline was captured. On the real MSRA117 poster, the metric change reflowed a research figure from 95.20% fill to 79.16%. The later validator only saw the already-shrunken layout.

The combined implementation maps Arial to bundled Arimo, validates migration before accepting a new authoritative baseline, and keeps all of #39's later durable and pixel-equality gates. It also preserves completed legacy logo bakes whose original auto-track width can no longer be recovered safely.

Validation

All tests below were rerun on this combined branch. Test data and results stayed under /tmp and are not included in the PR.

• all 6 modified Python modules compile; fit_logos.py --selftest, changed CLI help paths, git diff --check, and skill-creator validation for paper2poster and paper2reel pass
• MSRA117: four research figures render at 95%+; warm/cold 5760x3456 screenshots have 0 differing decoded-RGBA pixels; verify-final passes; repeated legacy and new-format rerenders produce byte-identical PNGs
• RSR-X1, RSR-X2, and Design-as-Code: research figures remain at 93% to 97%; independent cold-open HTML and PNG have 0 RGBA differences; repeated renders remain stable
• P01 legacy and local-first Arimo migrations: warm/cold and repeated-render comparisons have 0 RGBA differences; PDF decoded pages are identical; verify-final passes. The historical 79.898% figure is preserved only with the explicit legacy minimum-floor token
• negative fault injection maps Arial back to DejaVu Sans: the gate rejects the MSRA117 regression 95.20% -> 79.16% with exit 2 and rolls back HTML, PDF, PNG, and font assets unchanged
• additional low-fill real posters at 81.52%, 79.09%, 75.44%, and 83.06% are rejected rather than destructively backfilled
• full DS-1000 Reel bundle rebuild succeeds with 10 sections, 10 slides, and 10 timeline clips; package browser checks report 0 errors and 0 warnings
• Reel HTTP versus direct file:// screenshots are byte-identical at 1365x768, 1440x900, 1920x1080, and 2560x1440, with 0 differing pixels; repeated warm captures also have 0 differences; HTTP Range returns 206

Design-as-Code's page dimensions and page count are correct, but its pre-existing 24.26 MB PDF remains above the checker's 20 MB size cap.

PR hygiene

Exactly 12 production skill/runtime paths are included, matching the union of #39 and #40. No test files, screenshots, logs, generated posters, or audit results are included.