Bounded report preservation gate contradicts its own enrichment contract, failing every compliant final report #918

What happened

With #886/#892/#901/#909 in, smoke run 32898410734 (candidate a695b0a) executed the full pipeline for the first time — 2.5 hours of real agent work, all three workflows launched, every review stage ran — and all three targets failed at the very last node, final-report, on the report-severity-classification-preservation semantic gate:

Bounded report did not preserve dedupe field "lifecycle" at $.issues[0..3].lifecycle       (venus: every issue)
Bounded report did not preserve dedupe field "proof_of_concept"/"description" …            (very-liquid, stableswap)
Bounded report did not preserve dedupe field "triage_classification" at $.non_production…  (stableswap)

Root cause (provable from the contract itself)

The smoke lane's packaged topology omits severity classification, so the gate takes the bounded classification branch (reportBoundedDedupePreservationIssues). The final-report prompt for this mode instructs the agent to:

• "enrich its matching dedupe lifecycle record only in the report object" (triage_classification, triage_reason, final_disposition, demotion_reason, canonical_severity), and
• "choose exactly one triage_classification".

The gate honors that contract in its dedicated checks — lifecycleRecordPreservationIssues validates the report lifecycle against the authenticated ledger with BOUNDED_REPORT_LIFECYCLE_OWNED_FIELDS, and the row-level check requires row.triage_classification to equal the enriched lifecycle classification — but its generic loop then deep-equals every field of the dedupe finding against the report row, skipping only id/title for promoted rows. So whenever a dedupe row itself carries lifecycle or a preliminary triage_classification (schema-legal, and exactly what the venus and stableswap dedupe artifacts contained), a report that follows the prompt can never pass: venus failed on lifecycle for all four issues while preserving every prose field byte-for-byte — a maximally compliant report, rejected.

The gate landed in #558 (Aug 13) and was tightened in #621 (Aug 22), both after the last complete benchmark run (Aug 12), so the contradiction was never exercised end-to-end. Fifth latent post-Aug-12 regression surfaced by re-enabling the lane (#883, #889, #899, #908).

Fix

In the bounded branch's generic preservation loop, skip exactly the two fields that already have dedicated enrichment-aware validation: lifecycle and triage_classification. Everything else — including description, proof_of_concept, summary, evidence fields — keeps strict byte-preservation (venus demonstrates models comply). Regression test added: a dedupe row that embeds lifecycle + a preliminary classification passes with a compliant enriched report row, and tampering any non-owned field (e.g. summary) still fails.

The very-liquid/stableswap prose mismatches may additionally reflect model copy drift across retries; with the deterministic contradiction removed, retries can actually converge, and any residual prose flakiness will now be observable in isolation.