head: parallel batched bound walk + per-request k (the insitu ledger's first catch)

[physics515]

The in-situ ledger's first production run (post-#42, the 27B on the box) measured the bounded head at 469 ms/call vs the dense head's 68, evaluating ~91% of the vocab — the serial tile walk threw away the dense path's 16-way parallelism, and k=1024's threshold sits too deep in the bulk to prune this model.

Fixes, both measured-cause-first:

• Parallel batched bound walk: tiles evaluate across the rayon pool in batches; the threshold advances between batches; the stop rule is unchanged so exactness is preserved (overshoot bounded by one batch). Batch size ramps from ceil(k/TILE) to 64 so small-k pruning engages immediately and parallel width returns within a few doublings.
• Per-request k: serve scopes RequestTopK per request — 1 for greedy (temperature 0; the sampler reads only the argmax, and at k=1 the threshold is the running max, where the Cauchy–Schwarz bound actually prunes), the request's own top_k otherwise. Overlapping requests combine via fetch_max; the guard's drop falls back to the safe env default, so a concurrent request can never see a k smaller than its need.

391 lib tests green. The serve-side A/B on the 27B follows in the benchmark run this PR unblocks.

🤖 Generated with Claude Code