Run PR checks by affected components and preserve the required gate

[SaladDay]

Every PR previously ran the complete Core gate, including database tests and browsers for installer-only changes. Select checks from one conservative path map and the verified merge-parent diff, while keeping main and release checks complete. Unknown inputs or unavailable diffs select all checks.

The required check includes selected official-client and native acceptance and rejects failed, cancelled, missing or unexpectedly skipped jobs. Generated catalog and configuration references retain freshness checks. Distribution checks need no browser dependencies. Node jobs share a pinned pnpm store cache. Native keeps all three platforms, at most two concurrently, and still builds archives on PRs; successful uploads are limited to manual packaging and release. Release reuses the same native build once.

Validation: 25 planner/gate/metrics tests, 15 naming tests, 19 documentation/bundle tests, actionlint and complete Git-diff replay of 18 recent PRs. The replay includes deletions, renames, shared inputs and PRs exceeding 100 files. Fresh independent reviews covered the complete change. Substantiated findings were fixed and regression-tested; the last bounded fixture-mapping correction was locally validated without another review round. After incorporating the #314 test fix, the final full run passed all 15 jobs on head 6f32668b1f51f3185ebc06ba0afe2dfde8830bb1.

Measured Actions samples (raw machine time, not billed cost):

Case	Runner minutes	Wall minutes	Initial queue	Peak jobs	Result
Docs	0.50	1.52	10 s	1	Passed
Installer	1.48	1.77	10 s	2	Passed
Web	13.13	8.90	9 s	2	Passed, both browser shards
Full, final passing head	42.45	11.02	9 s	10	All 15 jobs passed
Full, before #314	40.62	12.18	80 s	10	13 passed; backend failed and gate correctly failed
Core, attempt 1	11.85	10.23	10 s	3	Existing database timeout/ownership failures; gate correctly failed

The Core probe’s failed-only rerun used another 9.47 runner minutes (9.85 wall minutes, 14 seconds initial queue, peak 1). The same inventory-retirement timeout failed; three prior successful jobs were reused and excluded from rerun machine time. The probe was closed without changing business tests or weakening the gate.

The prior successful installer runs consumed 23.12 and 27.17 runner minutes. The new installer sample used 93.6–94.6% less machine time. This is a small sample, not a long-term failure-rate or fee estimate. Full runs include API and three-platform jobs that the old core-check total excluded, so their totals are not directly comparable. Partial rerun statistics exclude carried-forward results and report them separately.

The earlier backend blocker was resolved by #314, which corrected the retirement test's competing deadlines without weakening production timeouts. The full CI run above verifies the integrated fix. Main's later #318–320 changes touch only disjoint Core files; all their paths remain covered by the existing backend/API rules. The user authorized administrator merge specifically for this PR after successful CI.

On the final passing full run, API and image acceptance, both browser shards and all three native platforms passed. Each platform created its archive successfully; successful PR uploads were skipped and the run retained zero artifacts. Platform runner minutes were Linux 35.10, macOS 4.25 and Windows 3.10. The required gate passed. Native bootstrap/recovery checks added by #310 also passed on all three platforms. Manual/release artifact wiring was reviewed and linted; no release was published to exercise it.

No business code, deployment, release publication or branch-protection change is included. Main impact selection and scheduled full runs remain out of scope.