perf(runtime): keep cattle conversations open across terminal runs

[Yevanchen]

Closes the V-C2 (warm-floor) direction of #387.

Problem

Production data (14d): 29/32 sessions receive their first message within 2s of creation — prewarm-on-create cannot beat the first-turn cold boot. The lever left is the follow-up turn, where the warm prepare path still paid ~4-5s because closeOnRunTerminal tore down the conversation session — and the driver process in it — after every cattle run. Per-run driver respawn also minted a fresh codex thread every turn, which scattered OpenAI's prompt_cache_key routing (codex keys it on its thread id; langgenius/mosoo-agent-driver#82).

Change

1. Cattle closeOnRunTerminal: false — the conversation session and its resident driver survive between turns inside the idle grace.
2. Conversation idle sweep (piggybacks the existing minute maintenance): closes cattle conversations quiet past the grace; the existing close path arms the subject inactive deadline, so container reclamation rides the unchanged subject-sweep chain. Conversations holding an active run lease are never treated as idle. Pet is untouched.
3. Cattle idle grace 90s → 5min — 90s covered only a third of observed follow-up gaps; 5min matches the read-think-type rhythm and the pet grace.
4. fix(api): route queue batches by name suffix — found by this gate: queue names are account-global, so isolated stacks deploy prefixed queues (mosoo-perf-stage-b-api-command); the consumer matched bare names only and silently acked every batch unprocessed. Scheduled maintenance had therefore never run on prefixed stacks (it stopped on the perf stacks the moment they switched to prefixed queues on 07-20). Suffix match; bare production names behave exactly as before.

Evidence (paired staging gate, openai instrument, stage-a=main 7aa09730 vs stage-b=this branch, same container image)

4 batches × 8 pairs — 32/32 pairs faster:

batch	main median TTFT	branch median TTFT	paired Δ	wins
1	7,861ms	3,132ms	−4,730ms	8/8
2	11,849ms	3,183ms	−8,569ms	8/8
3	10,586ms	3,156ms	−7,408ms	8/8
4	9,695ms	3,517ms	−6,086ms	8/8

• prepare_run warm median: 5.0–8.6s → ~1.0s (990–1,053ms across batches); branch-side batch medians vary ±50ms while main wobbles ±2s (driver-respawn lottery removed).
• Cache synergy confirmed: 4 sequential same-thread runs, cache_read = [10112, 12160, 12160, 12160] (4/4 near-full) vs [0, 0, 12160, 0] on main under the identical experiment — resident driver ⇒ stable codex thread ⇒ stable prompt_cache_key.
• Leak proof end-to-end (after the queue fix revived staging maintenance): probe conversation closed by the sweep → subject deadline armed → subject reclaimed and row removed; all stray conversations converged to zero. No schema changes, no migrations.

Cost ceiling: one extra ≤5min of container residency per session after its last run (the container already lives through the old 90s grace; the delta is grace length only).

Verification: api tc green, 1,053 api tests pass (incl. new sweep filter test), fmt/lint green.