[shiny-code-app]

Repeated native PostgreSQL jobs reach their 20-minute deadline without a failing assertion and block the merge train. Batch the invariant inspector's columns, indexes, primary keys and check constraints into fresh per-verification catalog reads. All existing invariant checks and write-fence refusals remain in place. The same-store ABBA comparison reduces SQL statements from 239 to 20 and median verification time from 0.246 to 0.031 seconds; a representative recovery scenario falls from 53.0 to 33.4 seconds.

Use seven partitions in postgres_integration through its existing generic discovery and partitioning. With batching alone, train candidate #4750's slowest native job still took 16:49. Replaying its measured 965 test costs projects worst test execution from 15:20 at six partitions to 12:19 at seven; eight projects 12:17, so seven is the smallest useful adjustment. Hosted performance at seven remains to be measured. All tests remain included exactly once and the 20-minute guard and 15-minute advisory stay unchanged. No other CI job changes. One additional runner job adds setup overhead; registered logical capacity is not proof of physical CPU capacity.

Validation: whole affected safety/schema/store modules (234 tests), native refusals (15), the original ordered 161-case shard, and affected native/schema compatibility modules passed against this runtime implementation. The latter selected 849 including a proposed budget probe subsequently removed; no existing test was changed or removed. Seven-way discovery verifies all 965 current source IDs exactly once. The 50 CLI/sharding/runner/security tests, actionlint, Ruff, formatting, role-word checks and mypy pass. Inspection NOT RUN/UNKNOWN under the Supervisor's explicit no-IDE instruction. Full new-head CI and independent final review are recorded in PR comments before checked Ready.

Closes #4188. PGTIMEOUT-1024, merge-path repair with explicit PR-cap exemption. Original stalled #4722 sources #4700/#4707/#4711 already landed through replacement #4731; no duplicate requeue is needed. Timing findings are on #4188 for CIMEASURE-1024.

Removed: repeated per-table invariant catalog reflection; no existing tests or assertions.
Net lines: +36 (+46/-10 across three files).

[shiny-code-app]

Review result: weighed
Head: ef3e9e3

Own independent review of record: Anthropic claude-opus-5-5[1m], confirmed by the reviewer CLI, completed on final head ef3e9e3. No implementation defects found in either round. Reviewed schema metadata shapes, fresh-read behavior, all invariant/refusal classes, absent/retired relations, callers and existing native refusal coverage.

Acted on the first review's test-design finding by removing the proposed catalog-budget test. The temporary probe exposed the tracked baseline's227 catalog queries, but its schema-size threshold could fail after an intended retirement. Existing behavior tests plus measured ABBA evidence qualify this optimization without a new brittle gate. No existing test or assertion changed. Shortened the existing testing paragraph.

Final review notes declined as non-defects: empty metadata filters can cause a bounded whole-schema read on a failing/retired-relation path, but do not restore per-table query growth or relax any refusal; a speculative extra branch is unnecessary for this repair. The operational testing paragraph retains the existing fresh-read contract and is shorter, not mutable runtime authority. Existing native tests exercise the adapter against actual PostgreSQL type/index/predicate/key/fence drift. Hosted job duration proof remains mandatory before Ready.

Before/after evidence: same-store/service ABBA, six old and six new, invariant SQL 239 → 20, catalog227 → 8, median 0.245669 → 0.031282s. Same recovery scenario 53.016 → 33.370s; unchanged ordered hosted shard of 161 tests passes locally in 250.353s. 234 whole safety/schema/store tests and 15 native refusal checks pass. Ruff/format/role-word and 1762-file mypy pass. Broader native-module run and current-head CI remain pending. IDE inspection NOT RUN/UNKNOWN because this brief forbids launching or touching JetBrains apps.

[shiny-code-app]

Ready for the merge train

Head: ef3e9e3

[shiny-code-app]

Launchplane completed the merge-train step for this pull request.

• Repository: cbusillo/launchplane
• Base branch: main
• Pull request: Restore PostgreSQL CI headroom with batched reflection and smaller partitions #4737
• Controller action: land_batch
• Controller record: merge-train-batch-landing-plan-20261024T210525Z-a5d9cf9e8c2b6d83

Launchplane landed this pull request through protected batch PR #4759.

Launchplane manages this comment and will update it as the train moves.

[shiny-code-app]

Review result: weighed
Head: 55cb802

Independent final review weighed for 55cb802: Anthropic claude-opus-5-5[1m], confirmed by CLI, through model-review helper. This follows the fresh design assessment after the earlier two review rounds.

No implementation defects found. Reviewer verified fresh per-call metadata, unchanged schema-refusal helpers, SQLAlchemy reflection equivalence and full native inventory partitioning. Missing-table exception wording remains a fail-closed nondefect. Empty-filter reflection is a bounded nondefect and the outer store rejects missing tables first. An unprofiled sibling reflection path is outside this measured repair; no speculative change.

Acted on residual operational risk by requiring actual new-head source CI and train-candidate PostgreSQL timings before completion. Seven partitions redistribute existing work and add setup overhead; the replay is an estimate and physical contention remains unknown. Seven has essentially the same projected worst case as eight while using one fewer slot. The 20-minute deadline and 15-minute advisory stay unchanged, no tests/assertions are removed, and no other CI job changes. Existing (2,3,6) sharding tests test generic properties and were correctly left unchanged rather than pinning workflow text.

Local new-head actionlint, 50 CLI/sharding/runner/security tests, seven-way exact-once discovery, Ruff/format, role-word check and 1762-file mypy passed. Prior whole affected safety/native tests cover the unchanged runtime implementation. IDE inspection NOT RUN/UNKNOWN under this brief's explicit ban. Actual hosted timings remain pending.

[shiny-code-app]

Ready for the merge train

Head: 55cb802