perf(int-factor): accept shared-host divisor evidence

[kim-em]

The divisor Phase-4 gate assumed a quiet or dedicated CPU could eventually be found, so complete evidence was repeatedly blocked by ordinary shared-host activity. This change adopts a repository-wide shared-host policy: pin for placement, use trial-major or adjacent AB/BA collection, retain every completed run, and treat host telemetry as context rather than an admission threshold.

For Hex.Nat.divisors, the preregistered seven-trial squarefree ladder now passes the independently derived τ log₂ τ model across all 42 measurements. The retained artifact validates complete divisor arrays and checksums outside timing; the exact LeanBench verdict is consistent_with_declared_complexity with residual slope -0.014231. The core and bridge reports map the transported APIs explicitly to runDivisors, clear Concerns, and recertify both libraries through Phase 7 without adding them to the release manifest.

The collector also preserves raw context and failures before validation, auto-selects the least-active available CPU without rejecting a busy host, and keeps reviewed runtime-neutral comparator changes from forcing unrelated remeasurement. Independent Opus review prompted removal of the remaining host-admission, preflight, retry, and obsolete null-control machinery.

Validation:

• lake build HexIntFactor HexIntFactorMathlib HexConformance HexManual (11,480 jobs)
• lake exe hexintfactor_bench verify (41 registrations)
• fresh fixture equals committed fixture; PARI checks 415 cases with 0 failures
• 226 collector, telemetry, freshness, and proof-sweep regression tests
• Phase 4, DAG, Phase 7, Mathlib-free, persistent-comparator, source-freshness, and figure checks

Closes #9619.