Fix poster HTML and Reel pixel fidelity

[Happy (huangyangyu)]

Summary

• freeze browser-selected Latin/CJK fonts into compact self-contained assets so standalone HTML keeps the measured glyph geometry
• persist fitted/expanded poster geometry across warm reloads and brand-new browser contexts, with a zero-difference decoded-RGBA gate before export
• keep Reel poster copies and inline srcdoc aligned for HTTP and direct file:// use, and bundle the Reel wordmark as a runtime asset

Validation

• skill-creator quick_validate.py passes for paper2poster and paper2reel
• all five changed Python modules compile; fit_logos.py --selftest and both changed CLIs pass
• real poster render with --mathjax-timeout-ms 0 produced PDF/PNG and passed the warm/cold 5760x3456 gate with 0 differing decoded RGBA pixels
• Trending backfill dry-run reports all 55/55 published papers already_current
• the public 2608.10299 Reel at 2016x1210 is decoded-RGBA identical to its canonical PNG: 0 / 2,439,360 changed pixels; its research figures clear the 0.90 floor (minimum 0.937805)
• fresh cross-week W31/W32/W33 sample: 12 papers x root poster / nested poster / HTTP Reel / direct file:// Reel = 48/48 screenshots raw-byte identical to their canonical PNGs; decoded RGBA changed_pixels=0, RMSE=0, 2016x1210, DPR 1

Scope

Exactly seven runtime files are included. No test folder or test files are part of this PR.

The production web backfill driver is intentionally outside this skill PR. Historical visible pixel identity is supplied by the canonical PNG layer materialized by that backfill; the renderer/font/geometry changes here harden future generated HTML and Reel bundles against the same divergence.

Legacy canonical posters that predate the current 0.90 research-figure floor remain immutable in this fidelity backfill. The backfill preserves their exact published PNG/PDF pixels instead of silently relayouting them; bringing those old posters up to the newer quality floor requires a separate regeneration that publishes new canonical PNG/PDF outputs.