[KaiEureka]

Adds logical model routes and shared request admission for independently started OC runs. A route resolves to an explicit physical model, provider and endpoint before capability checks. Processes on one host can share an existing request-pool directory; multi-host deployments use their configured common model gateway.

Request slots cover actual provider work and cancellation cleanup. Waiting for a slot is excluded from the generation timeout, while real generation deadlines and late-usage ownership remain intact. Team entry-role routing, CLI credentials, physical-model fallbacks and secret-free configuration errors match the SDK behavior.

Validation includes four independent Team/Workflow processes using two loopback HTTP backends with a shared capacity of two, a capacity-one parent/child run, queue cancellation, delayed cleanup and native retries. Independent review reproduced and repaired queue timeout, CLI routing and credential-diagnostic defects. The integrated isolation/Docker/model tree passed 67 focused tests. The model review also passed 95 timeout/admission cases and 180 CLI cases. Legacy duck-typed model clients retain optional-field compatibility. Ruff, import-linter and deptry pass. The final merge review passed 46 routing/admission tests and 543 existing compatibility regressions. The updated head includes both reviewed Docker import fixes. All final-head GitHub checks pass, including the complete Python 3.10–3.14 suites.

This is the third part of the runtime feature series, building on #242. Managed member-authored runs follow in #244.