ci: reduce selection fanout and local runner queue pressure

[jonit-dev]

Shared develop changes currently run all 13 template scenario legs and both golden journeys. Use starter as the default smoke, union explicitly changed runtime/scenario kits, and retain exhaustive main qualification, scheduled/manual full audits and unknown-input fallback. Template instructions retain scaffold/unit/instruction contracts without adding runtime kits.

Classify reviewed package producers and independent feature inputs in the existing integration selector. Shared core renderer/game-loop, patches and build configuration retain fluid coverage. Guard the independent modules/packages against the actual fluid fixture and verifier import graph; unknown executable dependencies and unclassified future producer consumers still fail closed. Required-plan recomputation, rename/deletion endpoints and merge-group unions remain intact. Move the web-reference and Android V8 source prerequisites to hosted Ubuntu 24.04 so local burst queues do not delay Android startup. Candidate identity, artifact transfer and consumer dependencies remain unchanged.

The exact published path-union model reduces template/golden legs from 15 to 2 for #381, #388, #395, #398 and #428; #416 retains 12 affected runtime kits plus one golden journey (13 legs). Integration source coverage narrows from 12 lanes to 9/9/3/9/8/9 respectively. PR-specific workflow deltas require fresh validation during rollout. #428 retains fluid because its core build configuration changes.

Allocate the local heavy pool as three slots with three whole physical cores each, using actual socket/core topology and keeping SMT siblings together. Three light join slots share a separate physical core; two physical cores remain reserved for the desktop. The read-only ci-runners.sh plan validates topology, available affinity and capacity before credentials or runner mutations. This halves heavy job concurrency while preserving its total CPU allocation; activation and measured throughput comparison follow a drained restart after merge.

Validation: fresh combined independent review, full CI contracts (376 passed), isolated workspace build and full workspace typecheck, Bash syntax/read-only actual-host allocation, and normal pre-commit/pre-push hooks. Regression tests cover changed template fanout, dependency selection, hosted producer contracts, whole-core launch shapes, malformed topology, capacity and restricted affinity. Live CI qualification is pending this followup. Previous integration run failed only the unchanged fluid capture on TN_PLAYTEST_OPERATION_TIMEOUT; runtime repair remains a separate owner. No assertions, timeouts, GPU access, security or registration settings changed.