refactor: unify media queue lane admission accounting

[atomantic]

Summary

Use the existing lane limits and classifier to account for GPU, cloud, and remote admission with one numeric slot map. Preserve snapshot timing, held-job skips, FIFO dispatch, live cloud limits, run-now oversubscription, and the worker cadence while reducing drainLoop complexity from 17 to 7 under the issue's counting convention.

Test plan

• Added a deterministic mixed-lane saturation/refill regression and verified it passes before the refactor.
• After refactoring, all 130 tests pass across mediaJobQueue/index.test.js, mediaJobQueue/sanitizeJob.test.js, and routes/mediaJobs.test.js.
• The regression exercises held local work, terminal-event refill, negative and zero cloud slots, per-lane FIFO, and live limit changes with isolated fixtures and generator stubs.

Closes #6677

[atomantic]

Implementation and local validation are complete: 130 tests pass across the three requested media queue/route suites, including the mixed-lane regression verified before and after the refactor. Claude ran locally at medium effort; its test-stub documentation/restoration note was addressed, while future-only lane concerns and the deliberately removed saturated-queue shortcut do not change this issue's chosen scope.

This PR is intentionally left open because required CI Gate remains red after one unchanged-head rerun. Attempt 1 failed the existing mediaJobQueue failure-hold reset test. Attempt 2 failed services/voice/tts-qwen3.test.js on Windows with QWEN3_RUNTIME_UNAVAILABLE and canceled sibling jobs. No CI bypass or merge was attempted. Branch and issue claim markers are preserved; the worktree is being removed for handoff.

[atomantic]

Review round 1 — lmstudio (qwen/qwen3-4b-2507)

Two findings, both resting on the same missing-context assumption; neither reproduces against the file.

"Blocking" — risk that laneLimits() returns a zero/uninitialised GPU limit, so GPU jobs would be skipped entirely. Not reachable: laneLimits() (server/services/mediaJobQueue/index.js:316) returns a literal gpu: 1; it is not configurable and has no zero state. The same helper already backs getQueueCapacity() and laneConcurrencyFor() on main, so this refactor reads the limit from the one definition rather than re-deriving it. The related implicit worry — a lane key with no slots entry making slots[lane] <= 0 compare undefined and admit unconditionally — is also closed: mediaJobExecutionLane() (server/lib/generationModes.js:67) returns exactly remote | cloud | gpu, matching MEDIA_JOB_EXECUTION_LANES, and slots seeds all three.

"Recommended" — the test does not verify that setCodexParallelLimit updates cloud capacity and is reflected in the running counts. It does: the final segment calls setCodexParallelLimit(2), ticks the worker, and asserts occupancy() moves to [1, 2, remoteLimit] with no worker restart. Negative and zero cloud slots (run-now oversubscription draining back through zero) are asserted earlier in the same test.

No code changes made. Local verification on this machine: 130/130 tests pass across mediaJobQueue/index.test.js, mediaJobQueue/sanitizeJob.test.js, routes/mediaJobs.test.js. The earlier CI run was cancelled mid-flight rather than failing; it has been re-run.