v0.443.0 — THE AUDIT BAND SHIP

B289: the corpus read in full and recomputed, the record of where the AI
tools failed, and a tender review — shipped as source. One band since
v0.442.0. No dispatch, registry-value, gate-assertion or physics change.

- audit/EXTERNAL_CONTACT_AUDIT.md (+ROWS.csv, rubric.txt, look_elsewhere.py,
  template_census.py, spotcheck.txt): 909 external-comparison lines
  classified; 4 pre-stated tests; 0 pre-stated-and-passed; the primitive
  vocabulary reaches any target within 0.1 pct (look-elsewhere).
- audit/THE_RECORD_WHERE_AI_FAILED.md: every real-valued constant traced to
  its birth chat and date (beta_i 0.603 mock-fit and [SSq] 0.57 wrong
  equation, both 14 Sep 2025); the 29-31 Mar 2026 audit-request -> G6-stamp
  fork; the retraction/recovery commits.
- audit/INSTRUCTION_FIDELITY_LEDGER.md + CONSTANT_DRIFT_TABLE.csv: 415 dated
  instructions vs what the tools did next, in the tools' own words; every
  value each constant took across 21 months.
- audit/WHITEPAPER_AUDIT.md + WHITEPAPER_AUDIT_ROWS.csv: all 2,328 papers
  read in full, headline recomputed: SOUND 13 / CALIBRATION_ONLY 878 /
  DEFECTIVE 547 / NARRATIVE 288 / UNSUPPORTED 265 / CIRCULAR 259 / STAMPED
  66; 0 of 25 pre-stated named comparisons hold; 448 GM/r^2-first.
- tender/CDG2752P27_REQUIREMENTS_MATRIX.md: OIL DOF tender mirrored clause
  by clause against uqff_downhole_simulator (program-improvement review).
- .gitignore: _to_delete/ (withdrawn a_DPM landmark parked, never ships).
- Registry: two audit-census rows (RECORDED). Arc marker AUDIT_ARC.
- Gate 6,115/0. Acceptance 109/109. Q-252 OPEN, Daniel-gated. Rule E held.
