docs: qualify historical Week 3 benchmarks after Week 2 rewrite

[skyzh]

Summary

• Label the Week 3 benchmark tables as historical results from the task #367 source commit 18aec850, before the five-day Week 2 rewrite. The old 43.1% 8K comparison uses the former Week 2 Split-K baseline; it does not compare against today's selected checkpoint.
• Keep the recorded measurements and teaching mechanisms, while sending new benchmark runs to fresh user-owned JSON paths instead of the tracked historical corpus.
• Preserve the appendix's copyable Bash commands in rendered HTML. Give the two key Day 5 serving and 8K comparison tables explicit row and column headers plus the book's horizontal-scroll wrappers.

Verification

The final local head 7d09adb6 builds the 49-page book and current sitemap, resolves generated local links and fragments, preserves the historical table values, validates the shell commands and rendered appendix code text, and places both wide tables inside the book's scroll wrappers. Independent exact-head reviews passed: Oracle factual (#861), Herald static accessibility (#860), and Scholar learner path (#862). A model benchmark was not rerun, so this PR makes no new current-performance claim.

Scope

Six book pages only. The remaining inherited benchmark-table row-header audit is tracked separately in task #855. This change does not publish a release or authorize a speed claim.