Harden EXP-001 preflight privacy and pricing

[sneakocom]

Purpose

Fixes exactly the two LOW findings from the independent review of merged PR
#15:

1. apply one privacy assertion to the explicit four-file public evidence set;
2. remove duplicated frozen input/output price literals and compute pricing
   drift from the frozen subject configuration.

This does not broaden EXP-001 preparation or change the frozen experimental
contract.

Privacy hardening

• Defines the intended public evidence set explicitly:
  model-catalogue.json, pricing-preflight.json,
  qualification-report.json, and value-receipt.json.
• Applies one shared forbidden-field/private-value assertion to every output.
• Rejects incomplete or unexpected public evidence sets.
• Adds synthetic-sentinel regression coverage for forbidden keys and values in
  each output; no private labels or authorization IDs are embedded in tests.

Pricing hardening

• Uses the existing frozen input/output prices in subject-config.json as the
  authoritative frozen spend assumptions.
• Computes price_drift_from_frozen_configuration from the catalogue and that
  configuration rather than hardcoding false.
• Rejects input/output drift against the frozen configuration and continues to
  reject cached-input drift against the current documented catalogue value.
• Preserves the conservative USD 6.25152 derived ceiling under the registered
  USD 6.30 limit.
• Account entitlement remains explicitly unverified.

Affected verification

Changed files:

• experiments/exp-001/live-preflight-v1/preflight.py
• tests/test_exp001_live_preflight.py

Selected locally because they directly cover every changed path:

• 11/11 dedicated live-preflight tests passed.
• Ruff passed on both changed Python files.
• git diff --check passed.
• Gitleaks passed on the changed source/test scope and on the exact commit.

The repository has 90 statically identified test methods. The other 79 were
not duplicated locally because no registry, dashboard, schema, captured
evidence, harness, coordinator, or shared validator changed, and repository
search found no other Python caller of the changed module. The normal PR
workflow may run its broader suite as repository-required verification rather
than impact-required verification.

Safety and claim boundary

• Provider/model subject runs: 0.
• Authorization consumptions: 0.
• Experiment runs/results: 0/0.
• Result envelopes: 0.
• EXP-001 remains PLANNED.
• No live provider request, account-entitlement probe, authorization mutation,
  private mapping access, or lifecycle advancement occurred.

This PR must receive a fresh independent review. Do not merge it as part of
this execution and do not launch EXP-001.