experiment pipeline: generate README tables from per-experiment fragments

[danReynolds]

Why

This completes the conflict-reduction work — the deferred README half of phase 2 (#202 did history/devices/signals). README.md was the last shared file experiments hand-edited, and the reason it was deferred: generate_history.dart parses README rows as a source, so generating it requires breaking that dependency carefully.

How

• experiments/index/NNN.json — one fragment per experiment with its README row ({file, title, impact, status, link}). Two experiments never touch the same file. A split experiment (014 has both an accepted and a rejected finding) stores an array of rows.
• experiments/README.template.md — the prose + section/table headers with {{ROWS:status}} placeholders.
• generate_readme.dart assembles README.md from template + fragments, newest-first. README.md becomes a generated, bot-owned aggregate.
• generate_history.dart now sources experiments from the fragments — it builds the equivalent table markdown in memory and feeds the unchanged, golden-tested parser (with a README.md fallback for fixtures). Its sort is made stable (insertion-order tiebreak) so ids sharing a number (008/008b, the two 014 rows) are deterministic regardless of source.
• check_generated_data builds README too; check_experiment_signals and finalize_experiment read the fragments instead of README rows; the bot regenerates README.md.

Correctness (round-trip proofs)

• README: regenerated from fragments, the row set per section is identical and prose/headers are byte-identical to today's README. The only change is row order, which normalizes to id-descending (the prior order was an irregular historical artifact — Accepted was recent-desc then old-ascending).
• history.json: byte-identical to main except the two duplicate-014 entries normalize to accepted-first (the old order was an unstable-sort artifact; consumers index by id, so it's cosmetic).
• Golden-fixture + pipeline tests pass; dart analyze --fatal-infos clean on all changed files; this PR touches no docs/*.json or signals.json (passes guard-generated-docs).

Rollout

Mirrors the signals.json bootstrap: this PR commits the reordered README.md as the new generated baseline, so experiments/README.md is not added to guard-generated-docs here — a one-line follow-up adds it once the baseline is on main.

Test plan

• check_generated_data — devices/history/signals/README all build cleanly
• check_experiment_signals — valid from fragments
• README round-trip: row set + prose preserved (content), only order normalized
• history.json round-trip: identical modulo the 014 stable-sort normalization
• dart test benchmark_pipeline_test benchmark_generated_outputs_test — pass (golden fixtures)
• dart analyze --fatal-infos — clean