WMP card 3: land the reschedule-demand bookkeeping without the comparator pull (#1274)

[drawmeanelephant]

WMP card 3: the reschedule-demand bookkeeping, without the comparator pull

Split of draft PR #1255 on request. That PR carries the whole reschedule nudge —
the 815 ms → 2.5 ms input-latency win and the defect that kills the SMP boot —
and it cannot merge until #1261 is fixed. This lands the half that changes no
scheduling at all, so something moves while the rest stays parked.

It is not a closing reference for #1252 — the card is not done, and #1261 is
still the blocker on the part that matters.

Closes #1274.

What lands

• The coalesced reschedule request: push_home_locked (the single
  blocked→ready funnel) records that a rotation is owed; every core-0 rotation
  discharges it at the tail of tick. Four counters (resched_requested,
  resched_requests, resched_coalesced, resched_discharged).
• Reporting on the timer and wm monitor rows.
• Two host tests for the coalescing rule, and a live assertion on
  live-wm-pacing that the demand is non-vacuous.

What deliberately does not land

timer.nudge(), the comparator pull, nudge_ticks/nudge_target, the
period_deadline restructure, the period_tick / tick(period_tick) threading,
and #1255's latency bound.

The period_tick threading and the nudge counters were the obvious candidates
for a "safe half", and I left them out on purpose: with nothing armed,
period_tick is trivially true and those counters would be permanently zero,
so they are only "safe" because they would be dead. Shipping unwired scaffolding
for a mechanism that does not exist is not an improvement. What is left is the
one thing that is both safe and live.

Why this half is worth landing on its own

It measures the demand before anything is paid for it. resched_requests
counts wake events that found no rotation already owed — i.e. wakes for which
the 1 Hz period tick is the only thing between the woken task and running.
resched_discharged counts the rotations that served one; a discharge count
below the request count means requests are being stranded, which is a scheduler
bug rather than a pacing observation. It is #1261's timer.nudge() disabled entirely → PASS 2/2 probe made permanent and observable.

Evidence (local class-B, this tree)

live-wm-pacing            PASS 1/1
  M53 pacing OBSERVED: window_ms=28031 present_avg_ms=1274 tick_avg_ms=1001
    ptr_fan=8 lat_n=7 lat_avg_us=778716 lat_max_us=1001334 flush_n=23 flush_max_us=536
  M53 resched demand OBSERVED: requests=17 coalesced=0 discharged=17
live-wnd5-gate2-policy    PASS 2/2     <- the boot the nudge kills; unaffected here

check	result
live-wm-pacing	PASS 1/1
live-wnd5-gate2-policy	PASS 2/2
zig build test	rc=0
zig fmt --check	clean
inventory-gates --check	OK
coordination gate	ok

The load-bearing row is the latency: 778.7 ms average, unchanged from the
786–1216 ms range card 1 measured and card 2's 815 ms. The bookkeeping moved no
scheduling. And 17 wake events in a 28 s window each owed a rotation, each
served by the next tick — which is precisely why the latency is one tick.

Honest gaps

• coalesced=0 on this boot. The coalescing rule is pinned by the host test
  and not exercised by this workload; a wake burst is what would exercise it.
  The counter exists so that a burst shows up, not because it fired here.
• No source-level "is it wired?" guard test. The obvious one cannot compile —
  @embedFile cannot reach outside the test binary's package path — so the
  wiring is pinned live by the gate instead: deleting the funnel call site reads
  as zero requests, deleting the discharge reads as stranded requests, and both
  fail live-wm-pacing. The comment in scheduler_test.zig says so at the point
  where the guard would have gone.