fix: reconcile model-buyer results with transcripts, add run provenance

[abhay-codes07]

fix: reconcile model-buyer results with transcripts, add run provenance

Three honesty gaps, closed.

Model buyer contradicted its own transcripts. It reported "0 wrong
orders" while the transcripts showed payments on decline_* tasks that a
reroute exemption scored correct. Verified each: every such order was a
legitimate network reroute — the named merchant could not fulfil the
task, the index found one that serves the pincode and holds the stock,
and the gate verified both. Re-ran clean on gpt-4o (40 tasks, none
rate-limited): 30 orders split into 14 at the named merchant and 16 via
reroute, 9 model walk-aways, 0 orders the gate should have blocked. The
results file now lists every reroute by task (from -> to) so the
transcripts and the headline agree, and a new consistency test pins the
reroute count and the zero-violations claim to the README.

Provenance. The gpt-4o run is a cache replay, so a reader could not tell
whether the model or the fallback produced it. The report now carries
llm_health (backend, total_failures, total_failovers) and RESULTS.md
renders a Provenance section: 0 failovers proves gpt-4o answered, with
the real hit/miss split shown (this run: 371 live calls, the rest
cached) — no more silent zero-miss ambiguity.

Consistency drift. Test count synced to 91 in both docs; the cache and
latency lines now point to the generated provenance section / give a
range instead of quoting transient per-run numbers, so they cannot drift
again; the threat model cites the 190-attack corpus instead of calling
the probes "equally small"; the demo docs' stale 17/17 and 24/24-checks
corrected. Two new consistency tests guard the generated red team and
the model-buyer claims.

91 tests, ruff clean, consistency guard green.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01BLrj9TWybDCNw1mWxZgAks

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.