The matrix inventories 69 required scenarios; it does not establish full PostgreSQL coverage of all 69. Each automated reference identifies a Cargo package/target, test function, required features, assertion scope, coverage layer and reviewed source SHA plus a checked LF-normalized test-file fingerprint. The current mappings are deliberately partial: a reference resolves to source, but last_run: null supplies no case-level run receipt. real_postgresql describes the test layer, not a successful execution. Counts are derived from individual records and checked by pg_matrix.
TC-060 now references event pagination in pg_read, which does not establish snapshot-to-delta continuity. TC-069 references shared admission-function labels, not three live transports. Authenticated Team HTTP/MCP transport and domain command dispatch remain unimplemented. These two cases do not contribute to the implementation count; the other implementation flags remain historical claims, not proof of complete coverage. Every mapping includes its evidence gap.
Historical real_agent_accepted records (42 cases) are preserved independently of automated coverage. Their structured references resolve through the sealed team-v1-historical-agent-evidence-v1.json index, which binds each case definition to its retained summary fingerprint. All 42 are historical_pending_review; current_verified is zero. Source and driver versions bound to the run, run identity and time, participant identities, explicit oracle expectations and an independent oracle identity remain unknown. Summary hash agreement locates retained bytes but does not revalidate the reported business result. Eight summaries also reference temporary outputs that were no longer retained; the index identifies those cases.
This correction did not rerun or revalidate those Agent sessions. The remaining 27 records are automated_evidence_pending, not an assertion that they cannot be replayed. The v1 historical index is sealed: adding or promoting an acceptance requires a new versioned evidence contract with complete bindings. The matrix guard checks references and accounting, not business acceptance.
A live dual-client run used Kimi Code CLI 2.0.0 and ZCode CLI 0.16.5 on one dedicated Team project (p11-live / work-p11): claim, implementation, handoff, verification, human review, completion. An independent oracle checked one receipt, the successor holder, and a single execution. That run is recorded as live_agent_run, separately from the later per-case records. The versioned matrix guard checks this summary independently: it requires the Kimi-to-ZCode client order, two distinct non-empty actors, the successor as active holder, exactly one receipt and one execution, a successful oracle, complete locators, and explicit release limits. It accepts new well-formed identifiers, client versions and Git SHAs; it does not treat the historical literal values as the only valid run.
A later read-only inspection observed the retained evidence bundle at 2026-09-21T14:33:31.256628+00:00. Its SHA-256 was 5257cc18c8e0ff13ed683abe569cefb65318c7361e728c8a1252c9beab4a17ec; the currently retained driver had SHA-256 81fa132d0d5ae61fcaaaed39a8f4021649d0b9702c83adc870dd7c7b3c0025d0. The bundle's oracle, receipt, execution and Git-head metadata agreed with the public summary, and its three embedded artifact byte hashes matched. The retained driver is not immutably bound to the historical execution, no explicit run timestamp is available, and the full historical chain was not reverified.
This correction did not rerun either live client or query author-private logs. Summary consistency is computed by the Rust pg_matrix guard; current_verification.live_rerun only records that no live rerun occurred in this correction. Neither the matrix nor its structural guard authorizes a release tag.
Capacity and reconnect probes
pg_capacity uses the shared fixture: only AWR_TEAM_TEST_DATABASE_URL is read, only loopback targets are allowed, and a newly allocated database belongs to this test process. It never reads the runtime AWR_TEAM_DATABASE_URL. Use disposable PostgreSQL instances; the fixture creates databases and an application role.
The capacity metric is combined_workflows_per_second: each of 20 serial workflows inserts a work item, starts a session and claims it. It is not isolated claim latency, enterprise throughput or a production SLA. Setup and version lookup are outside the timed interval. Successful output is visible with:
cargo test -p awr-team-pg --features pg-tests --test pg_capacity -- --nocapture
To persist JSON, set AWR_CAPACITY_REPORT to a new file, AWR_TEST_SOURCE_SHA to the measured 40-character commit, and AWR_TEST_SOURCE_DIRTY to true or false. Preserve a diff alongside a dirty measurement. The report records the scope, elapsed seconds, rate, PostgreSQL version, OS/architecture and timestamp; it excludes connection URLs and credentials and refuses to overwrite a report. Without a report path, stdout/stderr must be retained explicitly.
pg_fault separates healthy reconnection from incompatible-version and missing migration-record controls. The latter assert schema rejection and unchanged work-item bytes in the process-owned database; they do not simulate every network outage or claim production HA.
