fix(spec): re-sync the #102-moved prose figures and bind them in verify:claims

[justin13888]

Summary

The #102 re-score moved spec/EXPERIMENTS.md §7.13, §7.10 and §11.9, but four prose figures that restate those cells by hand kept their old values, and verify:claims bound none of them.

• spec/V0.8-DECISIONS.md — D1's prior evidence: LOO coder saving −7.4% → −7.5% (§7.13, per-index context backing off to order-0, LOO / vs fixed).
• spec/README.md — CfL roadmap row: free −0.09% → −0.10%, paid +2.18% → +2.11% (§7.10). Scalefactor-bands roadmap row and the §15 summary paragraph: −0.13% → −0.14% (§11.9).
• tools/comparison/src/verify-claims.ts — six new register entries binding those quotes (plus the CfL row's −0.90% code-3 figure), and spec/V0.8-DECISIONS.md added to the header's list of quoting files. The register now checks 34 figures across 6 files; registering spec/V0.8-DECISIONS.md also puts it under the unregistered-figure scan, which finds nothing there.

Validation

• node tools/comparison/src/verify-claims.ts — exit 0: "Checked 34 quoted figure(s) across 6 files", no unregistered figure.
• Falsifier: with the four old values temporarily restored in the two spec files (not committed), the same command exits 1 and reports all five stale quotes (7.4 vs 7.50, 0.09 vs 0.10, +2.18 vs 2.11, 0.13 vs 0.14 twice).
• pnpm --prefix tools/comparison run format:check — pass.
• pnpm --prefix tools/comparison run lint — pass.
• pnpm --prefix tools/comparison exec tsc --noEmit — exit 0.

Coverage gaps

• ci-comparison.yml's path filter does not list spec/V0.8-DECISIONS.md, so a PR editing only that file does not run verify:claims (a re-score that edits EXPERIMENTS.md still does). Outside this change's scope; filed as fix(comparison): run verify:claims on spec/V0.8-DECISIONS.md edits, and re-sync EXPERIMENTS.md §11.13's scalefactor cell #115.
• spec/EXPERIMENTS.md:2911 (§11.13) still says −0.13% for scalefactor bands. That is the source document, which this gate cannot check; also in fix(comparison): run verify:claims on spec/V0.8-DECISIONS.md edits, and re-sync EXPERIMENTS.md §11.13's scalefactor cell #115.

Risks and rollout

None. Documentation figures and a CI documentation gate only; no format, API or encoder change.

Decisions taken

• Where §11.9's best arm is bound. §11.9 states −0.14% only in prose and has no table, and the register binds cells. Taken: bind both quotes to §9.5's re-source table, row Scalefactor bands, best arm, column Wikimedia, which holds the same measurement (−0.14%). Rejected: extending the register to bind prose (a new mechanism, and it would gate a sentence rather than a cell); adding a table to §11.9 (outside the manifest, and EXPERIMENTS.md edits belong with the sweep bindings). Reverse: change section/row/column on the two scalefactor-band entries.
• How the paid CfL figure is bound. +2.11% is against the shipped row, not a table cell. Taken: bind the CfL paid, L24@5 C9@4 / ΔE00 cell and recompute against shipped / ΔE00 in a transform, the same shape as §11.14's WebP margin; the capture keeps the + and omits abs, like the 108 B entry, because the sentence's content is the direction. Rejected: abs, which would pass a flipped sign. Reverse: edit that entry's transform/pattern.
• Binding the −0.90% code-3 figure too. Not named in fix(spec): re-sync prose figures the #108 re-score moved, and bind them in verify:claims #113, but it is the other end of the same sentence and the same column. Taken: register it. Reverse: delete the CfL with free gains at code 3 entry.
• CI path filter and the §11.13 cell left out. Both are outside the manifest (.github/workflows/ci-comparison.yml, spec/EXPERIMENTS.md). Taken: filed as fix(comparison): run verify:claims on spec/V0.8-DECISIONS.md edits, and re-sync EXPERIMENTS.md §11.13's scalefactor cell #115 rather than widening. Reverse: fold fix(comparison): run verify:claims on spec/V0.8-DECISIONS.md edits, and re-sync EXPERIMENTS.md §11.13's scalefactor cell #115 into this branch.

Issue

Closes #113