QA run · integration-system (6/14 items consulted) · e4e5c6e3 · 2026-08-17 · 2 PASS / 4 PARTIAL / 0 FAIL / 8 NOT-RUN #9338

R3 runner, area 4 of 4. Part of the #9296 wave. No clause failed. Eight items were not reached and are recorded not-run rather than inferred — including the area's new P0, which is called out separately below.

⚠️ Title deviates from the wave's (FULL area) convention on purpose — 6 of 14 items consulted.

Environment fingerprint

Same tree as #9330 / #9332 / #9337: framework e4e5c6e3c608b1b807c83a0d5b734f213eb1a1dd · .objectui-sha 665661ab093263f39f2e660a295ea615dbcee35a · console dist stamp equals pin (check:console-sha ✓) · in-process VerifyStack boots for the dogfood pins, package-local vitest for the rest · 2026-08-17.

Spec provenance: checklist from PR #9309 head fbd1c049 (open, held pending the maintainer), not merged main. Validator (pinned): 190 items (190 active); 30 kinds mapped, 0 waived. Selector area:integration-system → 14 runnable, 0 blocked.

Automated pins (RUNNER rule 6)

Per-clause verdicts

integration-system.connector-stdio-default-deny — rev 2, P1, cli → pass

Pin mcp-provider.test.ts — 13/13. This is the item's sole declared automated.ref and it covers the default-deny posture the item exists for. No unpinned clause remained, so pass rather than partial.

integration-system.connector-descriptor-audit — rev 2, P2, cli → pass

Pin connector-descriptor-audit.test.ts — 9/9, against 3 acceptance / 2 negative clauses. Sole declared ref, fully green.

integration-system.connector-declarative-boot — rev 2, P1, api → partial

Pin showcase-declarative-mcp.dogfood.test.ts — 2/2:

• "materializes all three generic-executor instances at boot — mcp included, state ready"
• "dispatches the MCP tool end-to-end through the flow connector_action"

Covers boot materialization and end-to-end dispatch; the remaining acceptance clauses (6 total) were not separately driven → partial. The item's stale-dist trap is structurally ruled out: every package was rebuilt from source in this container before any test ran.

integration-system.webhook-lifecycle — rev 4, P1, mixed → partial

Pin webhook-materialization.dogfood.test.ts — 1/1: "materializes the stack-authored webhook into a sys_webhook row (object→object_name, isActive→active)". That is the authoring→row half only; delivery, retry and the console half of this 6-clause item were not driven → partial.

integration-system.email-template-render — rev 4, P2, mixed → partial

Pin email-template-materialization.dogfood.test.ts — 2/2:

• "materializes the stack-authored template into a sys_email_template row (bodyHtml→body_html)"
• "renders the authored template through sendTemplate — the disconnect #4488 审计发现的四个"授权门断连"：email_template / job / validation 的元数据条目到不了执行点，action 导航项点不动 #4509 closed"

2 of 8 acceptance clauses' worth of surface → partial.

integration-system.datasource-admin-lifecycle — rev 1, P1, api → partial

Pins admin-routes.test.ts + envelope.conformance.test.ts — 54/54 passed, pinning route behaviour and envelope shape. partial is the item's own stated ceiling, not a shortfall of this run: its automated field says the pin "pins route behavior + envelope; the LIVE-mount half is not pinned, drive os dev for it". The live-mount half was not driven, so the dispatcher-vs-hono-route trap this item carries remains open — the pins run against in-process dispatch, which is exactly what that trap warns cannot substitute for a real socket.

⚠️ integration-system.datasource-credential-refusal-matrix — rev 1, P0, mixed → NOT-RUN

New in the R4 sweep, never run, and not reached by this run. Recording it plainly rather than partially: nothing was consulted, so there is no verdict.

This is the most consequential gap across all four of my areas and should be picked up first. Why it needs care, from reading the spec (not from running it):

• Its clauses are leak checks. A0 requires one publish attempt per inline-credential door enumerated from packages/spec/src/migrations/entries/semantic/; A2 requires grepping the full response body of both read doors for planted cleartext; A7 repeats both reads as non-admin and anonymous. A cleartext credential in any read response is N0 — a P0 leak.
• If a leak is found, it must not be filed publicly. Per the wave's standing rule, the run issue marks the clause fail with the detail withheld, says so explicitly, and the full reproduction goes to the maintainer in a comment on tracking: full platform-checklist regression against main @ e4e5c6e3 — five-round orchestration #9296. This item is the single likeliest place in my four areas to trip that rule.
• Two documented false-fail traps: encryptionKey (turso) is deliberately still writable while being redacted on read ([services half of #7990] Close the datasource/connector credential write/read paths: scrub getDatasource().config, fix the false "credential-stripped" claim, and write the stored-cleartext-rows migration story #8081 item 4) — scoring it as a write refusal is a false fail; and the legacy-alias clause (A4) tests a row shape the current parse refuses to create, reachable only by planting a stored row directly. If it cannot be planted, A4 is blocked(fixture) and must be recorded — the item warns that an untested redaction is worse than an untested feature.

Not-run (7 others)

No oracle consulted, not inferred: connector-degraded-recovery (P1) · connector-spec-path-no-escape (P2, build) · flow-connector-picker (P2, browser) · job-scheduled-run (P1) · notify-inbox-delivery (P1) · external-datasource-federated-read (P1) · notification-preference-suppression (P2).

One incidental reading relevant to external-datasource-federated-read, from the objectstack verify sweep in #9330: showcase_ext_customer and showcase_ext_order were skipped by the RLS prover with the reason "external read-only object (federated datasource "showcase_external"; no inserts)" — confirming the federated fixture is wired and identifiable. That is context for the next runner, not a verdict on the item.

Coverage honesty

6 of 14 items consulted; no clause failed; 2 pass, 4 partial, 8 not-run. All evidence is server truth (test oracle) from declared automated.ref pins — no clause in this area was hand-driven against a live socket, which is why every mixed item stayed partial and why datasource-admin-lifecycle explicitly retains its live-mount gap.

Links: #9296 (wave) · #9330 (area 1) · #9332 (area 2) · #9337 (area 3) · #9309 (spec provenance) · #8081 (turso encryptionKey scope).