Release v0.3.5: simjoin CPU hot-path speedup (accumulate + verify), bit-identical

[prostomarkeloff]

What

Speeds up the simjoin CPU join hot path, byte-for-byte identical output. Bumps the crate to v0.3.5.

Profiling the real find-dup-defs patternology join (n≈3216 functions, dense bi/tri-gram vectors, θ=0.85) with samply phase-split + Apple-Silicon PMU counters showed the bottleneck there is not cos_full (that's the sparse PyPI-287k regime) but the ~5M candidate touches in accumulate and their per-candidate bound check — throughput was held down by branch mispredicts, not memory.

Changes (all output-preserving)

• accumulate — split the acc/touched borrows so the acc base pointer is no longer reloaded from the Scratch struct on every posting entry; and make first-touch branchless (unconditional store into touched, committed only by bumping the length). Same accumulator values + same touched order ⇒ bit-identical. Removes the data-dependent mispredict that capped IPC (3.0 → 5.1, cond-mispred 5.8 → 1.2 /1k insns).
• verify_pruned / collect_survivors / cosine_join_counts — a cheap monotone pre-bound: since xpn is monotonic, xpn[kstar] ≤ ‖probe‖, so a candidate failing a + ‖probe‖·pnorm < need also fails the exact L2AP bound — prune it without the per-candidate partition_point. Survivor set is bit-identical. Gated on di.len() >= PREBOUND_MIN_DIMS so the sparse regime (short rows, cheap binary search, high survivor rate) pays nothing.
• examples/simjoin_pypi: add SJ_NSUB env to bench a row subset (small-n regime).

Results (byte-for-byte unchanged)

workload	speedup
find-dup-defs patternology join (cycles 309M → 180M)	~1.4–1.55×
PyPI type3 287k join	~1.05× (no regression)

Correctness

Parity is the crate's hard gate — all green:

• indexed_join_matches_bruteforce (O(n²) oracle, 400 fuzzed corpora)
• gpu_hybrid_matches_cpu (Metal f32-filter / CPU-f64-reverify)
• Both real corpora: emitted pair sets identical (1849 / 4 427 097)
• cargo clippy strict (all + pedantic, incl. gpu) clean; full test suite green