[criptogus]

Summary

Agent Match gains optional consent-bound Jev relevance selection, source-bound generation and bounded streaming. The v6 candidate changes address repeated selector calls and unstable filtered prefixes while preserving the original relevance thresholds, full-catalog recovery, generator profile and all approval criteria.

approved current catalog + explicit consent/config
  exact credential-scoped probability cache (fixed 30 min / 64 entries)
    hit → current catalog + unchanged coverage rules, zero new selector usage
    miss → all named candidates, compact shared policy, validated usage
  fresh generator call → strict schema + literal source evidence

Cache keys bind all semantic inputs, the current eligible catalog, model/protocol and provider credential fingerprint. Anonymous requests bypass reuse. No raw brief, credential, generated answer or approval is cached. Benchmark scope additionally binds the cohort; orphan/expired cache evidence cannot qualify. The server rejects caller balances above the actual accumulated allowance.

Evidence

• Before: complete frozen v26 (72/72) passed all six quality dimensions, zero high/critical, but failed efficiency: conservative total baseline US$0.8725 vs Jev US$1.623483414; client p95 23,805 vs 28,230 ms. Ten of twelve identical-input cases produced different subsets across all three repeats.
  After: actual adapter regression changes two transport calls to one for exact repeated input; invalidation, consent, credential isolation, expiry, capacity and failure tests pass. This is deterministic mechanism evidence, not a new cost/p95 result.
• Before: client remaining=US$7 could bypass previously billed spending in the server reservation ceiling.
  After: the red/green regression rejects resetting the remaining allowance. Prior billed upper is US$5.48125, unknown holds US$1.08116575, remaining US$0.43758425 within the unchanged US$7 cap.
• Cache-window regression: all 24 repeated selections in the original unchanged sequential schedule have intervening-call lower bounds of 371,917–447,318 ms, already above the prior five-minute TTL. The seven-minute clock-controlled regression failed with two calls and passed with one after fixed thirty-minute expiry; a read does not extend lifetime, and the test proves expiration at thirty minutes. Distributed/cold workers may still miss. This changes no original quality, cost or p95 criterion.
• Compact serialization reduced request bytes 54.54–54.80% in all twelve frozen cases. One real five-candidate TypeSafe JSON-instruction contract probe returned actual jev-1.14.0, 990/89 tokens. Neither byte counts nor a selector-only contract probe establish generator quality, savings or p95.
• 285 tests and all eleven local project checks pass; CI for the new commit is pending; strict lint reports zero blocking errors and zero explicit-any violations. New scripts additionally passed standalone strict TypeScript; actual diagnostic rejection probes cover report overwrite, budget reset and unbound cache provenance.
• The v28 plan 95d12543a45a6dad14ebad083ae378f5d74714d95fe7c41aad068a7778a7c52b preserves cases/catalog/desired/schedule/generator/criteria exactly, with fresh runtime hashes. The empty-evidence report was correctly rejected with zero observations and approved:false. The new 72-generation comparison was not started: US$0.43758425 cannot fund the US$2.38 reference cohort; even observed reference output-only provider cost is US$1.245384, a planning reference rather than a guaranteed future bill. No additional Lovable inference, builder usage or publication. No partial cohort or recycled generator output is counted as approval.

Merge Danger

Door: two-way.

Blast Radius: matching.

Do not merge/deploy or activate filtering on this evidence. v26 remains rejected and v6 is unqualified. Keep this PR draft. The temporary preview benchmark endpoint must be removed before any separately qualified production release. No business-skill versions, approval/compatibility records or RLS changed. Full methodology and actual local evidence locations are documented in docs/AGENT-MATCH-PAIRED-EVALUATION.md and the ADR.