Integrated candidate and CI followups

Draft PR183 integrates the Discovery stack onto main at 0d5fa2c0c. Its initial commit 46facd017 has tree 568d6368a25bec46288561f2cfbb47c462ff9a83, exactly matching the ordinary merge at d96f0a69a. Existing stacked PRs and their history remain intact. No scanner was disabled, incident dismissed, shared history rewritten, or main branch pushed.

Initial integrated checks

• Hosted build: five successful tasks; typecheck: eight successful tasks.
• Thirty-two focused latest-main generation/operator/billing tests passed.
• GitGuardian check 106053058648 passed on exact head 46facd017.
• Discovery CI passed the live ClickHouse job and hosted build/typecheck, then failed one worker deadline regression in the server selection. Subsequent verify-job steps did not run; the overall workflow was not green.
• The separate full OSS export workflow passed on that same head. Service-dependent skips remain explicit in its logs.

The historical stacked-PR GitGuardian findings pointed to fixture values in b796ef73e, already replaced in current files. The new ordinary integration candidate received its own unchanged scanner check; this does not dismiss the historical incidents or change the old PRs' check results.

Deadline regression and fix

The existing test expected stopReason: time but received failed. The worker selected the smaller of its provider timeout and remaining run budget, but its timer always produced a generic provider error. A timer firing just before the wall clock reached the deadline lost the reason that selected the timer.

The timer now preserves its selected bound: run deadlines and equal limits report time, shorter provider timeouts remain nonretryable failures, and external cancellation retains its behavior. Request/unit charging, retry limits, cost journaling and publication checks are unchanged. No timeout was increased and no assertion was relaxed.

Frozen-wall-clock/real-timer regressions reproduced the original failure, then passed after the fix. The final worker, cost, transport and journal selection passed 52 tests, zero failures or skips, including synchronous retryable rejection during abort and late embedding output. Independent review found no blocking issue, passed 39 worker/transport tests and exercised timeout/abort ordering separately. Its suggested synchronous-abort case was added permanently.

Selection optimization

Commit 279fbd2ed removes unused full-card fields/counter calculations during queue selection while preserving final delivery checks. Thirty-five local regressions and an independent 31-test review selection passed. The two matched Railway arms each passed 96 HTTP checks and 1,890 SQL calls without errors; all temporary resources were removed. See the projection and measurement report for exact evidence and why the observed latency difference is not a causal claim.

The complete updated source also passed hosted build and typecheck (five/eight tasks). All GitHub checks subsequently passed on exact followup head a70a3aa87e3ed8c355978708e8ced087f2a2a74e:

• Discovery CI: hosted build/typecheck; 776 server tests passed with eight explicit service-dependent skips; 24 Redis capacity tests; 91 client tests; 185 Python dataset/evaluation tests; client/export boundaries. The separate live ClickHouse job passed all nine tests, including the warehouse regression skipped in the general server job.
• Full OSS export passed: shared 55 passing/one skip, app 786 passing/no skips, server 1,489 passing/four skips. These selections overlap other checks; do not sum them into a unique-test claim.
• GitGuardian and Cloudflare Pages checks passed on the same head.

The timer failure on the initial integration remains part of the evidence; it is resolved by the preserved-deadline fix, not by changing its assertion.

Remaining gates

Production flags were reread after these tests: cursor, measurement, personalization, treatment percentage and interest onboarding are all unset; their defaults remain off/zero. Dedicated Discover Redis is not configured.

The existing ClickPipe was last verified Running with 17 table mappings and still lacks discovery_events and discovery_erased_actors. Source and warehouse schema preparation does not establish real CDC receipt or erasure coverage. Full hosted entrypoint/proxy/multiple-instance testing also remains distinct from the isolated four-route origin probes. The complete hosted deployment followup subsequently used a new owned Redis service and two app services. It exposed and fixed client-clock observation loss, then passed all seven phases and 73 HTTP checks on e5322c710, including continued paging after one app service was removed. All temporary resources were independently checked absent. Production Neon capacity and controlled production rollout remain separate gates.

The subsequent clock-fix revision e5322c710 also passed complete hosted CI, including 102 client regressions, live ClickHouse and OSS export. Its full deployment report preserves the initial packaging failure and the clock-skew failure rather than relabeling those attempts as passes.

Experimental representation quality still failed its frozen promotion test. Preserve the baseline and require fresh evidence before changing ranking. No production deployment, model promotion, paid provider run, or new subscription occurred in this validation followup. The overall rollout goal remains open.