feat(bb/msm): coop-walker — workgroup-cooperative inversion bucket accumulator

[AztecBot]

Re-architecture: cooperative-inversion bucket accumulator (coop-walker)

The stream-walker accumulate kernel is memory-bound / occupancy-limited; its
occupancy is capped by two costs that scale with S (slots per thread): ~150+
live registers and a 16 KB var<workgroup> pref_scratch (one resident
workgroup on Mali). coop-walker sets slots-per-thread=1 and shares the
batched inversion across the whole workgroup (interleaved prefix/suffix product
scan + one safegcd inversion per round): per-thread registers ~8× lower,
workgroup memory 16 KB → ~4 KB, ~8× fewer inversions, still streaming each
point once. Drop-in for the stream_walker dispatch (accum: 'coop').
Details: COOP_WALKER_DESIGN.md.

Correctness — GPU vs @noble/curves, headless SwiftShader

coop GREEN (gpu == noble): logn 8/10/12/14, multiple seeds. Walker too.

Real-hardware timing (BrowserStack, min ms over reps)

Apple M2 (macOS, Chrome 148) — desktop-class, NOT memory-starved:

logN	walker	coop	coop speedup
12	15.0	14.6	1.03×
14	23.7	43.2	0.55×
16	49.4	107	0.46×

coop loses at logn≥14 on M2. Expected: M2 has ample shared memory + many
cores, so the walker is not occupancy-starved there, and coop's structural cost
(1 add/thread/round ⇒ ~8× more rounds, each with workgroup barriers) is not
repaid by latency-hiding it doesn't need. The design's premise — the 16 KB
scratch capping the walker to one resident workgroup — only bites on
mobile (Mali 16 KB), which is the regime the measured ground truth
describes. Mobile (Adreno/Mali) runs are in flight to test exactly this.

SwiftShader agrees directionally (coop wins logn≤12, loses at 16) but is
correctness-only — it models none of the occupancy/latency-hiding the design
targets, so neither it nor M2 is the design's intended regime.

Status

• ✅ coop-walker kernel, correct, drop-in, pushed.
• ✅ harness: WASM-free GPU-vs-noble + same-device multi-logN A/B sweep, real
  devices despite the 213-byte wasm stub.
• ✅ Apple M2 real-HW numbers (above).
• ⏳ Adreno (S25 Ultra) + Mali (Pixel 9 Pro XL) — the memory-starved regime.

Reporting mobile numbers as they land. If coop wins on mobile but not desktop,
that is the honest, useful result: the structural win is occupancy on
memory-starved GPUs, and it does not help GPUs that already have occupancy.