docs: maintainer pass — consume feedback backlog, reconcile GPU dead-ends + stale numbers

[ChrisLundquist]

Maintainer pass consuming the .claude/feedback/ backlog into CLAUDE.md. Docs-only.

Promoted (CLAUDE.md edits)

• GPU dead-ends reconciled with PR feat(pz2): G32 stage 2 — Metal literal decode 65 GB/s, but the splice is the wall (GPU decode killed with data) #154's pz2-G32 stage-2 data. The old "GPU entropy (rANS/FSE) is slower than CPU — 0.54x decode" was a serial-state topology artifact: a wire-codesigned Metal kernel decodes literals at 64.75 GB/s. GPU block decode still loses, but on the LZ splice / shared-memory ceiling (0.21x at the real operating point, 1.02x parity at 16x synthetic saturation), not on entropy. Encode-vs-decode framing now split.
• Per-core decode wall entry amended: the wall was format-structural, not hardware. pz2 (sequence wire + 8-lane Huffman + fused splice) decodes ~3.5x faster per core at ratio parity — replace the format, don't micro-opt lzf. Profiling gotcha preserved.
• pz bw decode row 1120 → 1367 MB/s (PR perf(bw): K=8 multi-cursor inverse BWT — 1.55x ST / 2.66x all-cores CLI decode #151 K=8 multi-cursor inverse BWT, 2.66x all-cores); multi-cursor rider + GPU-iBWT rejection documented.
• Num row: -a now auto-routes via numeric_gain (feat(analysis): auto-route fixed-stride numeric data to the Num pipeline #138); added the "don't trial-encode a 64KB sample" gotcha.
• New dead-end: vertical bit-packing can't replace per-plane FSE in the Num pipeline (+18/+10/+2/+47pp).
• Key conventions: vec![0;n]/alloc_zeroed vs with_capacity+resize on fresh decode buffers.
• Test caveat: test_lz77_batched_pipeline_not_slower_than_serial flakes under multi-agent load (verified: passes in isolation).

Discarded / already-covered (notes deleted)

• 6 March-2026 notes (GPU hash tables, GPU rANS, GPU LZ-win/entropy-lose, LzSeq family, per-stream overhead, ratio-gap-is-encoding) — all already covered by current dead-ends / architecture sections.
• ci-clippy note — the --all-targets fix already landed in its own PR.
• numeric-decorr-frontend spike note — superseded by the shipped Num pipeline.

All claims verified against docs/design-docs/pz2-g32-stage2-findings.md, ibwt-cursor-findings.md, num-bitpack-stage0-findings.md, and clean-slate-codec.md.

CI note: the one local test failure is the documented flaky wall-clock GPU perf assertion (758/758 real tests pass; flake passes in isolation at 0.21s). Change is docs-only.

🤖 Generated with Claude Code