fix(models): benchmark the observed model through its installed inference route

[tang-vu]

Why this matters

The Models page could show the loaded model while Benchmark discarded its installed route's TLS, API prefix and authentication. It could also treat a saved GGUF_FILE as a live loaded identity after the runtime reported no loaded model or its health endpoint failed, then send a completion request that implicitly loaded that stale model.

Benchmark now uses the existing destination-bound chat route resolver and requires an observed loaded identity before sending inference. Direct runtimes receive that observed identity; the shipped LiteLLM gateway receives ods/current. Missing runtime evidence returns 503 without a completion request or saved performance sample. After the owner loads a model, a fresh request can benchmark it normally.

Validation on 37510d0b4d110c8d14280c98eb31b5c5310af73e, based on current upstream main 49c1e6e376ce1771cbe9572426de1d0a2ca0477d, native Windows and Python 3.14:

• The new authenticated API regression uses a real local HTTP runtime, a real installed-artifact fixture and the production catalog builder. Null loaded health and unavailable health both failed before the follow-up fix: the endpoint returned 200 after sending an unwanted completion for the persisted filename. Both now refuse before POST, preserve the artifact and recover after live health reports a loaded model.
• Benchmark transport, request measurement and shared setup route suites: 49 passed. Includes the existing real authenticated HTTP endpoint, TLS/prefix routing, destination-bound credentials, invalid-route refusal, gateway aliases and the new refusal/recovery cases.
• Related models and performance oracle suites: 209 passed at the final SHA.
• Python compilation and Git diff checks passed.

The existing PR history is preserved; current main was merged without conflicts. Only routers/models.py and tests/test_benchmark_transport.py are changed against main. #2471 validates malformed model-observation shapes; this change does not duplicate its observer hunks. #6992 repairs registered-store activation idempotence independently.

The shipped Lemonade v10.2.0 server specification documents that chat completions auto-load a model that is not already loaded. The HTTP fixture verifies ODS request behavior; it is not a live Lemonade or GPU performance claim.

No fresh Docker/WSL installation, live Lemonade process, model inference or browser end-to-end validation ran. Concurrent replacement after a valid initial observation remains an existing limitation. Request usage/wall-time measurements retain their explicit label and include gateway and transport overhead.

Overlap check

Searched open and closed upstream PRs for benchmark transport, TLS/prefix authentication, observed model identity and no-loaded-model behavior; inspected changes to routers/models.py. #2471 handles malformed observation shapes and is not duplicated. #6992 covers registered-store reloads, and #7213 covers Hugging Face credentials. This updates the existing canonical #6990 rather than opening a competing PR.

AI Assistance

Codex implemented the original route fix and this observed-model safety follow-up, reproduced the failure and conducted an independent final-head review. Human review remains required before merge.

Release Lane

• Mainline change targeting main
• Stable hotfix targeting release/2.6.x

Changed Surface

• Dashboard API / host agent
• Model routing / Hermes / capabilities
• Network exposure / auth / proxy

Risk And Validation

Medium risk. The exact-SHA focused and related receipts above passed; independent reviewer reran 13 transport tests without blocking findings. git diff --check passed. No data migration is introduced.

Operational Change Check

• This is an operational change and validation is intentionally deferred for live model, GPU, Portal and fresh-install execution. Real local HTTP tests qualify ODS transport/refusal behavior and are not release-grade fleet evidence.

Combined batch validation

These five independently useful PRs were merged without conflicts onto synthetic integration head 0f8d40fa2116a0ea4111d1aff4905f8a9c07a549, based on main 49c1e6e376ce1771cbe9572426de1d0a2ca0477d: #7212, #7213, #6990, #3835, #7214. This records the tested order, not a merge dependency.

Combined native validation: 339 dashboard API/models/performance/workflows/agent tests passed, 3 live n8n-key tests passed, and 8 native PowerShell 5.1 CLI HTTP tests passed. git diff --check and the PowerShell parser passed. The HTTP/SDK fixture and unavailable live Docker/model/Portal limits stated above still apply; integration success does not close those gates.