perf(coro): streamline demand-free syscall handoffs

[cpunion]

This reduces the remaining scheduler-side cost of compiler-certified native syscalls without weakening blocking compensation.

What changes:

• Arm deferred syscall compensation without eagerly allocating a replacement directory slot. Only a durable request that wins Armed -> Starting materializes the slot and physical owner.
• Consume the compiler-hidden task capability to freeze the current driver route and avoid repeated complete source-catalog/owner-local audits during detach, quota release, and reentry.
• Keep the original full validation path for TLS/reentry, retained, and other non-compiler callers.
• Add race/unit/source-architecture gates for the slotless Armed/Starting protocol and compiler-task detach path.

Measured on Darwin arm64, Go 1.26.5, LLVM 22.1.8, full LTO, GOMAXPROCS=1, 15 rotated process runs:

workload	scalar parent	candidate	delta	candidate / Go
standard file, 5,000 ops	42.343 ms	38.376 ms	-9.37%	4.94x
direct syscall file, 5,000 ops	18.485 ms	15.826 ms	-14.39%	2.07x
loopback TCP, 500 ops	19.980 ms	19.343 ms	-3.19% (overlapping ranges)	2.04x

Size changes: stripped binary +1,520 bytes (+0.022%); Mach-O __text +3,700 bytes (+0.118%).

Validated locally:

• full runtime module
• coroutine architecture-debt and native target-plan gates
• same-M scheduler progress, locked compensation, and request-driven direct-channel replacement E2Es
• real standard file, direct syscall, sole-M blocking pipe + timer, and loopback TCP executions

The next performance target remains synchronous child-frame allocation/publication/destroy through syscall, internal/poll, os.File, and io; this PR deliberately does not mix that lifecycle work into the native handoff protocol.