Stop the served findings contradicting the evaluation

[jadoon200]

Verifying the live deploy caught the shipped Q13 finding still asserting that daily
autocorrelation "predicted every outcome measured so far" — the exact claim the Q13
experiment refuted hours earlier. The verdict string, the explainer and the docs were all
updated; the Finding record baked into the snapshot was not. The deployed site was making a
claim the project had already disproved.

Fixed

• Q13 → refuted, with the serverless numbers, the non-monotone shape, the failed
  spectral-entropy replacement, and the low-frequency lead that does not survive Bonferroni.
• Q4 quoted "19 of 24 settings", which no configuration of evaluate_frontier.py can
  produce — it emits three traces × six ratios. Replaced with what the corrected run gives:
  strictly better in 9 of 18, worse in 6, identical in 3 at the ratio where p95 and the
  derived target coincide by construction. The regime matters more than the count and is now
  stated: derived wins below 19:1, loses above.
• Q10's evidence was one sentence with no numbers. It now carries the frontier figures.

Tested, not remembered

This is the third time a served string has drifted from docs/EVAL.md — after the
verdict text claiming nothing below 0.50 ever won, and the explainer claiming an unsupported
scope. Prose can't be fully pinned, but the specific disproved assertions can.

test_findings.py pins that Q13 does not reassert the disproved claim, that Q1 names the
percentile recommender rather than threshold HPA as what it lost to, that Q4 does not quote
the unreproducible count, and that no confirmed finding claims a win for forecasting without
naming the regime it holds in. It also catches thin evidence — which is what flagged Q10.

Live deploy verified

delphi-c5pp.onrender.com is up and serving from the current main: /health reports
snapshot_mode: demo, all three band states render, the interactive diagnostic computes
end-to-end (r=0.724 → verdict, q*=0.9500 from prices), /geoint/evidence is ARGUS-shaped,
zero console errors. README now links it.

Gate

ruff clean; mypy strict clean; 195 tests passing (was 190).