fix(research): RSI correction round — token matching, artifact-derived reports, lineage and provider validation

[rustfuture]

Why

A prior audit found the reported multi-seed numbers did not match results/multi_seed_results.json.
Root cause: README.md carried a hand-written results table that no committed artifact ever
produced; report.md was generated but nothing checked it, and the README was never regenerated.

Re-running the committed v1 code reproduces first_run.json, history.jsonl, metrics.csv
byte-for-byte and the multi-seed/ablation numbers modulo runtime_seconds, so the artifact was
sound and the prose was stale.

What changed

• Experiment definition (versioned). Keywords now match whole tokens, so the positive keyword
  safe no longer fires inside unsafe and clear inside unclear. experiment_version = v2-token-match, schema_version = 2. v1 artifacts preserved verbatim in results/archive-v1/.
• Reporting. rsi_framework.reporting renders results/report.md and the README results block
  from results/*.json. --check and test_report_numbers_match_artifacts fail on drift; CI runs
  the check.
• Counting. "Accepted versions including the baseline" is now reported separately from
  "accepted new changes".
• Provenance. results/provenance.json records commit, dirty paths, config hash, dataset hash,
  splits, run command and interpreter.
• Lineage. History rows carry a unique event_id distinct from the policy version. A repeated
  policy hash in a later generation no longer overwrites earlier records; missing parents are
  reported as orphans and cycles are detected. The recorded parent_event_id is always the accepted
  event that held the active policy.
• Provider injection. run_experiment(..., candidate_generator=...); JsonFileProposalProvider
  reads externally produced proposals; validate_proposal rejects no-ops, multi-edit proposals,
  illegal tokens, keywords outside the vocabulary and out-of-bounds bias. OpenWeightColabL4Stub
  is documented as an unwired boundary; no LLM was run.
• Ablations. Added ablation_random_selection (genuine score-free control); ablation_no_selection
  relabelled as dev-blind greedy selection; each mode records the mechanism it changes and runs with
  the same seed, dataset hash and budget.
• Terminology. rollback_regression → rejected_dev_regression (the rule rejects a candidate;
  no applied state is reverted). Legacy label retained as a v1 alias.
• Claim scope. Byte equality limited to the four runtime-free artifacts; leakage test described as
  a scoped invariance check; held-out n=8 stated wherever outcomes appear; multi-seed runs described
  as replays of one shared synthetic pool.

Verification

• python3 -m unittest discover -s tests -v → 23 tests, OK (also on python3.11).
• python3 -m rsi_framework.reporting --check → consistent.
• cmp on first_run.json, history.jsonl, metrics.csv, progress.svg after an independent run → byte-identical.
• Re-run benchmarks → multi_seed_results.json / ablation_results.json equal modulo runtime_seconds.

Pending

• Real open-weight LLM integration is not done. The injection point and validation gate are
  implemented and tested with a file-backed provider; the model evaluation is explicitly pending.