# TCID-13132 internal evidence fidelity and SQL text deduplication Initial base: `25165d8225ec1bda9b2122a7e6704c2bb4d92ec5`. Integrated base: `69f112606f8cf9f921b41249f09932c50a49c2ac`; preserves the independently merged Wetech WSS egress change. Historical source: isolated replay `40c29173-b31d-4d5e-b977-d91e11015426`, image source `65232e57`, answer fingerprint `6b64aab42876ab219675e57e32b7731e269c8029df5ab4d1642bb54d6c07d342`. Judge: `different_but_valid`; diagnostic completeness is a separate criterion. The query named `tiflash-peer-process-cpu-cores` returned five TiDB instances with `job=observability/tidb-pool`. Scenario Diagnostic interpreted them as TiDB, not TiFlash. The earliest missing evidence is in acquisition; historical Agent9 audit retained argument hashes rather than PromQL, so the exact selector cannot be reconstructed. Returning executed PromQL prevents that loss in future Provider results; it does not restore historical Agent9 audit payloads or change Agent9's generic audit implementation. The replay did not match the exact target SQL or retrieve its execution plans. Its broad Slow Query lists therefore cannot establish equal actual rows across the human's 4.5–20.8 second target executions. SQL deduplication preserves available facts but cannot manufacture that absent evidence. Recovery at 11:27 is outside the historical replay cutoff of 10:57:09 UTC+8 and must not leak into the replay answer. TiFlash source availability and exact-SQL acquisition remain separate diagnostic follow-up questions, not claims of this patch. Offline application to the original exported handoffs keeps all five currently retained facts in each scenario. Serialized UTF-8 sizes change from 115,052 to 93,191 bytes (query-latency), 116,092 to 94,231 (tiflash-io-apply), and 115,737 to 78,901 (workload-hotspot). These historical outputs contain short Digests, so only identical SQL text is coalesced. This check does not restore facts already removed by the original replay or predict the next full replay's evidence set. Immutable package artifacts were prepared and read back without installing or deploying them. Local regression runs no model or Provider and performs no Jira publication or optimization-status write. Keep both problems at `candidate` until an authorized deployed isolated replay verifies their corresponding boundaries. Validation: `npm test` passes with 1,204 passing tests, no failures and two browser-integration tests skipped. Coverage includes full returned Digests, executed PromQL, SQL/log preservation, credential redaction (including SQL authentication clauses), repeated SQL with preserved per-execution facts, cross-database/target isolation and immutable package/source matching. Final prepared versions: Clinic 31, Manus 7, Wetech 63. Bundle SHA-256: `3c9cd1c3d786bd339655304c55a2d95c472dbc4391dcbcad57995babeed339bd`. The package was read back by its exact object version and verified byte-for-byte; `deploy:agent9-tools -- --validate-only` passes without deployment. ## Subsequent deployment and verification The preparation-only statements above describe the initial stage. PR #1402 was subsequently deployed with source `363d4d0b` and the bundle above. Read replay `5f9cb4a4-ec38-45be-87a0-218614c36001` completed on 2026-09-07 18:41:43.420 UTC: Judge `different_but_valid`, acquisition complete, no final failure codes. The actual Clinic output now retains full Digests and executed PromQL, and the runner's cumulative handoff reports SQL deduplication. It still retains only 9/15 query-latency observations under the total budget; deduplication does not promise unlimited context. TiFlash metric emptiness remains a diagnostic gap. The target SQL was found after the separate runner-input fix recorded in `2026-09-08-tcid13132-supplied-sql-context.md`; that outcome cannot be attributed solely to this fidelity patch. No Jira/DB/Base refresh or Wetech/Manus functional case replay was performed, so these problem records remain unverified.