feat(bb/msm): per-architecture autotuner for stream-walker TPB (Mali/Adreno/Apple)

[AztecBot]

Summary

Makes the stream-walker faster (its slow path) by combining two correctness-neutral, memory-free levers on top of the per-architecture TPB autotuner:

1. dx-caching — the walker's batched-inversion inner loop touched each point's x-coordinate ~3× per add (forward prefix pass, inverse pass, backward peel), all uncoalesced SRS reads. The forward pass now caches each slot's dx = p_rx − p_lx in private registers; the inverse pass reuses it instead of reloading point_x and re-subtracting. This removes ~1/3 of the walker's uncoalesced x-coordinate reads at zero device/workgroup memory cost (S×8 = 64 u32 of registers/thread; the swarm established register pressure is not the walker's binding constraint).
2. Per-architecture TPB autotuner — picks the largest threads-per-block whose pref_scratch footprint fits the device's maxComputeWorkgroupStorageSize (TPB=128 on 32 KB Apple/Adreno, 64 on 16 KB Mali). TPB only repacks a fixed set of NUM_THREADS = nwg×256 logical threads into workgroups; each thread owns its own [local_id×S, +S) scratch slice with no cross-thread sharing and no barrier, so it is correctness-neutral.

Both compose: dx-caching cuts per-add memory traffic; the autotuner raises occupancy/packing per arch.

dx-caching (the speed lever)

pref_scratch[k] holds the forward prefix product; the inverse pass peels it
back to per-slot 1/dx. It needs the same dx the forward pass computed — caching
dx (bit-identical) lets the inverse pass skip its point_x reload + subtraction.

_generated/shaders.ts is regenerated (inlined WGSL) and included in the diff.

Device-memory pref_scratch — explored, gated (honest finding)

The swarm flagged moving pref_scratch to device storage as the #1 occupancy lever. Wiring it revealed a real hardware ceiling: the walker already binds all 10 storage buffers, so a device-memory pref_scratch needs an 11th. SwiftShader and many mobile GPUs cap maxStorageBuffersPerShaderStage at 10, which rejects the pipeline layout.

This PR adds the walkerPrefDevice knob + WGSL variant (binding 11, global t*S*2 slice) but gates it on maxStorageBuffersPerShaderStage ≥ 11, falling back to var<workgroup> otherwise so it never breaks. To engage it on the ≤10-buffer mobile targets a storage binding must first be freed (e.g. interleaving sorted_bucket_list+sorted_count_list, which share an index) — tracked as follow-up; not landed because it can't be perf-validated without a real-hardware seat.

Validation

• Autotuner unit tests (host, no GPU): 10/10 PASS.
• SwiftShader correctness (software WebGPU, GPU-less host) — cross-check vs @noble/curves BN254 Pippenger, GREEN at logn=8 and logn=10 for: workgroup pref_scratch at autotuned TPB=128, forced TPB=64, and the device-pref fallback path. SwiftShader reports maxStorageBuffersPerShaderStage=10, confirming the ceiling above.
• tsc --noEmit clean for msm_webgpu/.

Real-hardware A/B — still outstanding (seat contention)

The 2 shared BrowserStack seats have been continuously saturated by other agents' Mali (Pixel 9 Pro XL) and Adreno (Galaxy S25 Ultra) jobs across this session's patient backoff. Not run rather than displace other agents' work. The harness is ready (?tpb/?prefdev in the xcheck page; --qp passthrough in run-browserstack; cloudflared installed). When a seat frees, the dx-caching delta is a direct apples-to-apples comparison vs this PR's previously-recorded macOS M2 TPB=128 stream_walker accumulate of 33.1 ms (same code minus dx-caching):

# Apple (M2): dx-cache delta + TPB A/B
node dev/msm-webgpu/scripts/run-browserstack.mjs --page xcheck --target macos          --n 16 --reps 5 --qp "tpb=128"
node dev/msm-webgpu/scripts/run-browserstack.mjs --page xcheck --target macos          --n 16 --reps 5 --qp "tpb=64"
# Adreno (wall-clock; timestamp-queries unreliable) + Mali (TPB autotune differentiation)
node dev/msm-webgpu/scripts/run-browserstack.mjs --page xcheck --target s25-ultra      --n 16 --reps 5 --qp "tpb=128"
node dev/msm-webgpu/scripts/run-browserstack.mjs --page xcheck --target pixel-9-pro-xl --n 16 --reps 5 --qp "tpb=64"

Status

Draft. dx-caching + autotuner + gated device-pref knob landed and SwiftShader-GREEN. Mobile Apple/Adreno/Mali numbers are the only outstanding item, blocked purely on shared-seat contention.