perf: improve seckill consumption and timeout dispatch

[ChanTso]

Commerce could admit requests faster than the single-hotspot order consumer completed them, and long runs exposed a second backlog in timeout dispatch. This change permits four bounded order consumers, narrows inventory locking, and makes timeout dispatch proportional to the ready batch instead of the accumulated queue.

• Use a shared activity lock for read-only checks and conditional stock/version updates before inserting the order. Preserve one order per transaction, rollback on publication races, and ACK after commit.
• Register four actual receive/transaction/ACK schedules. Add concurrent-stock, publication-change and scheduler coverage.
• Index the generated ready-to-dispatch predicate in retry order. Send MQ messages outside the database transaction, then commit the batch's delivery receipts in a separate transaction; a lost receipt permits replay, and duplicate cancellation does not restore inventory twice.
• Use the same 10-second observer for both comparison versions and handle only the observed k6-exit EOF case. Retain raw outputs and all unsuccessful probes.

At the same local 200/s five-minute load, baseline and final code each completed 60,001 orders with no dropped iterations; SQL reservation-to-order p99 changed from 9.26s to 1.70s. The repeat completed all 59,963 admitted orders but had 37 unissued startup iterations. The 400/s run still accumulated orders and had one negative receiving-time sample, so it is an overload/recovery case, not a certified sustained capacity. All final SQL inventory/ledger invariants passed. The ready-query average changed from 15,004 examined rows / 47.85ms to 4.61 rows / 0.191ms in the approximately 320-second native sampling intervals.

Environment: local M4, Docker 8 CPU / 14GB, Commerce 4 CPU, MySQL buffer pool 1GiB, durable commit settings unchanged. No positive warmup. Historical fixtures and shared Broker timers are disclosed. Final measured code is 9bb08c6; the report commit is not substituted for that SHA.

Validation: make java-ci python-ci web-ci repo-ci passed; make test-catalog-integration passed 207 tests, including 17 seckill transaction tests. Independent read-only code and evidence reviews found no blocking issues. Repository checks and secret scan passed after the final reports. The final code's remote 26-check gate passed. The merged main commit e6bed53 also passed all 26 checks in https://github.com/ChanTso/citybuddy/actions/runs/34203315877. Original signing metadata and MySQL cache size were restored and Docker stopped; isolated fixtures remain explicitly UNPAID/SENT, not falsely recorded as cancelled.

Results: final comparison, transaction overview.