honest fix-forward (cont.): clear Codex confirm-pass residuals

- RELEASE_CHECKLIST: lead shipped 0.5B cross-source 0.980 (label 0.984 as 1.5B nested-CV
  upper bound); minpairs same-distribution: text slightly beats shipped probe (~0.74 vs ~0.82).
- README badge: "runs 100% local" -> "runs local" (prose already qualified).
- (docs/promo.html, a stale untracked marketing page, removed from disk — was never on GitHub.)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: ithiria894

README.md

@@ -10,7 +10,7 @@
   <img src="https://img.shields.io/badge/Python-3.10%2B-blue?logo=python&logoColor=white" alt="Python 3.10+" />
   <a href="LICENSE"><img src="https://img.shields.io/badge/License-Apache%202.0-blue" alt="License" /></a>
   <a href="https://doi.org/10.5281/zenodo.19990741"><img src="https://img.shields.io/badge/DOI-10.5281%2Fzenodo.19990741-blue" alt="DOI" /></a>
-  <img src="https://img.shields.io/badge/runs-100%25%20local-brightgreen" alt="Runs locally" />
+  <img src="https://img.shields.io/badge/runs-local-brightgreen" alt="Runs locally" />
   <img src="https://img.shields.io/badge/telemetry-zero-blue" alt="Zero telemetry" />
   <img src="https://img.shields.io/badge/status-research%20preview-orange" alt="Research preview" />
 </p>

docs/RELEASE_CHECKLIST.md

@@ -124,8 +124,10 @@ Safe to say:
   evidence spans, thresholds, policy reasons, and scanner artifact id.
 - its edge is generalization to unseen sources/wording: on held-out real attacks
   (HackAPrompt) it recalls 90.3% at a 5% clean false-positive rate vs a same-data text
-  classifier's 52.8%; curated cross-source AUROC 0.984 vs 0.914 (CI-backed).
-- on matched-vocabulary / same-distribution it TIES a text classifier (~0.79 vs ~0.82);
+  classifier's 52.8%; curated cross-source AUROC 0.980 vs 0.914 for the SHIPPED 0.5B config
+  (0.984 is a nested-CV upper bound that may pick a larger 1.5B sensor — label it as such).
+- on matched-vocabulary / same-distribution a text classifier slightly BEATS the shipped 0.5B
+  probe (~0.74 vs ~0.82);
   tool-poisoning generalization is PARTIAL and on SYNTHETIC attacks (MCPTox the one
   significant win). novel attack-family generalization is still the open frontier.