Phase 2 of the Tracker quiescence-handshake optimization. Phase 1 (the waiting-flag gate) takes per-cycle futex_wake from O(N) to O(1) by only notifying when the WaitSet thread is actually parked. This issue tracks the deferred phase 2: a bounded spin in wait_for_quiescence before parking, so that in the common case (jobs finish in microseconds) the dispatch thread never parks at all — waiting stays false, no futex_wait, no futex_wake, zero syscalls per cycle.

This was deliberately split out of phase 1 because it is the only piece carrying real-time scheduling risk and its entire justification is empirical. It must clear the acceptance criteria below before merging.

Goal

Eliminate the residual O(1) futex_wait (waiter parks) + futex_wake (last completer notifies) per cycle by spinning briefly on the submitted/completed counters before taking the lock and parking.

Acceptance criteria

• Hard gate — measured jitter win on hardware. Must demonstrate a dispatch-jitter improvement on the Pi5 via the existing CycleObservation histogram, A/B against phase-1 main. No hardware proof → no merge. (This is the Pi A/B that was declined for phase 1; for the spin it is mandatory, because the spin trades CPU for syscalls and its only justification is the measured result.)
• Safety — no priority inversion. Bounded spin only. Do not spin when the dispatch thread shares a core with a SCHED_FIFO worker: affinity and FIFO priority are caller-configurable per crates/taktora-executor/src/thread_attrs.rs, and a spinning higher-or-equal-priority thread on a shared core can prevent the very worker it is waiting on from being scheduled. Gate the spin on core topology (or skip it entirely when topology is unknown/shared).
• Tuning — no magic constant. SPIN_BOUND derived from measured job-completion latency on target hardware, not picked by feel.
• Loom model still passes (the spin path must not introduce a lost wakeup): during the spin waiting is false and completers correctly skip notify; only on spin exhaustion does the waiter set waiting + recheck + park.

Context

• Phase 1 handshake: crates/taktora-executor/src/pool.rs (Tracker::complete / wait_for_quiescence).
• The two-hazard ordering rationale (Dekker SeqCst flag-decision race + lock-protected park handshake) is documented in-code and enforced by the loom model — the spin must preserve both.