fix: falsification-audit hardening — exact conservation, -O guard, empty-index join (C-67/C-68/C-69)

[Polichinel]

Fixes all four findings of the 2026-07-02 four-axis audit's falsification phase (1 hard, 3 soft).
Bug fixes with an identical contract — the code now honors what the docs already promised. No
public-surface change; CONFORMANCE_FLOOR stays 1.0.0.

The hard finding — C-68

reconcile_proportional's torch-port + 1e-8 denominator epsilon was a float32 no-op for draw
sums ≳ 0.1 (why no oracle/parity test ever saw it) but silently deflated the scale factor for
tiny nonzero sums: at draw-sum 1e-8 a country total of 100 reconciled to 50, no error — violating
the Reconcile.md §3 sum-to-country guarantee in-domain; the package's own assert_reconcile_contract
rejected its own output. Fix: explicit all-zero-draw guard + exact division — conservation now
exact for any nonzero sum; all-zero draws stay zero exactly as before; bit-identical on all
realistic data (torch-oracle parity green, unchanged). Also: negative country totals now raise
ValueError instead of silently clamping to a zero output (F8).

The soft findings

• C-67: the three published conformance suites (bare asserts) silently no-op'd under
  python -O — a verified float64 non-conformer passed. Now: _require_assertions() on every
  entry point → RuntimeError under optimized bytecode (refuse-to-run beats report-green-while-dead).
• C-69: empty-index searchsorted crashed with an obscure IndexError (the clip(0,−1)
  corner) → now returns all -1, the documented not-found value.

Ritual

• Regression pins: tests/test_falsification_safety_audit_2026_07.py (written as failing stubs
  during the audit, turned green by these fixes; the F3 pin runs a python -O subprocess).
• Docs: Reconcile.md §6 + proportional.py module docstring (records the deliberate,
  parity-preserving deviation from the torch original). CHANGELOG [Unreleased] entry.
• Register: C-67/C-68/C-69 registered-and-resolved, C-70 opened (the audit's docs/tests polish
  bundle); counts 63→67 total / 13 open / 54 resolved. /review-rr triage CLEAN.
• Gate: ruff · mypy (3.10+3.11) · 398 tests, 100% line+branch · validate_docs — all green.
  /review-diff CLEAN.

🤖 Generated with Claude Code