docs(job-manager): restore GPU barrier protocol divergence

docs/dev/debt/libSceJobManager.md

@@ -2,7 +2,7 @@
 
 ### Functional
 
-- [sce::Job::JobManager](../../../core/libs/prx/libSceJobManager/Export.cpp) submits commands as work becomes runnable rather than encoding ahead with the rewind protocol recovered from PPSA06888's module (offsets 0x2210 and 0x2650). Commands are recycled after a completed wait. Command capacity is a separate bound from the dispatch count: the native size query measures one register and a 31-character marker, then its exhaustion callback at 0x29d0 traps instead of growing storage. The fallback throws. Original-module execution with the host AGC builders accepts two dispatches with 255 registers and 65,535-character markers, then traps on a third. Console timing equivalence remains unverified. Compute dispatch on a CPU sequence throws. Shutdown with outstanding compute work throws.
+- [Compute sequences](../../../core/libs/prx/libSceJobManager/Export.cpp) encode ahead across GPU barriers using explicit release packets instead of the native rewind protocol and completion-packet patching recovered from PPSA06888's module (offsets 0x2210 and 0x2650). Commands are recycled after a completed wait. Command capacity is a separate bound from the dispatch count: the native size query measures one register and a 31-character marker, then its exhaustion callback at 0x29d0 traps instead of growing storage. The fallback throws. Original-module execution with the host AGC builders accepts two dispatches with 255 registers and 65,535-character markers, then traps on a third. Console timing equivalence remains unverified. Compute dispatch on a CPU sequence throws. Shutdown with outstanding compute work throws.
 - [Scheduling](../../../core/libs/prx/libSceJobManager/Export.cpp) uses host mutexes and condition variables instead of native atomic queues and semaphores; spin intervals deliberately use unscaled host TSC, while the original module reads scaled guest TSC under the host-only `APS5_TIME_SCALE` diagnostic. Blocks below 0x2d8 bytes, factory priorities above one, use of destroyed sequences and shutdown from a worker throw. New calls concurrent with shutdown are unverified.
 - [Pool allocation](../../../core/libs/prx/libSceJobManager/Export.cpp) reuses freed nodes in a different order; profiler dispatch IDs can differ from native allocation. The fallback Wait deliberately allocates no pool node; native Wait can stall on global pool exhaustion. Resource-creation failures throw after cleanup, where the original does not check every return. Vtable order was checked, but no calling title or PS5 timing equivalence has been tested.
 - Console behavior for [worker waits](../../../core/libs/prx/libSceJobManager/Export.cpp) is unverified. The owned original module stalls on the zero-spin, zero-persistence `worker_wait` fixture with native Windows and Linux host kernel services; the fallback wakes a spare worker and completes the child. The [reference runner](../../../tools/run_job_manager_reference.py) on this branch supports Windows only; failures remain failures.