docs: GPU strategy doc and north-star plan update

[ChrisLundquist]

Summary

• Add docs/design-docs/gpu-strategy.md — a single document capturing the GPU compression strategy based on empirical results across the project's history
• Update PLAN-unified-scheduler-north-star.md to reflect current reality (Phase 1 done but perf gate failed, Phase 3 done via PR Unify GPU schedulers into a single coordinator thread #101)

What the strategy doc covers

• What GPU is good at: LZ77 cooperative-stitch matching (1,788 parallel probes, 94% quality)
• Why hash tables failed on GPU: atomic ordering destroys match quality (6.25% on repetitive data)
• Compression ratio gap: PZ-LZR 41% vs gzip 29% — primarily a match-finding quality gap, not entropy
• What GPU is bad at: entropy coding (rANS at 0.77x CPU encode, 0.54x decode — serial state machine bottleneck)
• Current architecture: unified scheduler with GPU coordinator, try_send() deadlock prevention, GPU-to-CPU fallback
• The FusedGpu problem: routing entropy to GPU is currently counterproductive
• What would need to change: match quality improvements, on-device chaining (blocked on GPU entropy parity)

North-star plan changes

• Removed stale "Critical gap: No GPU rANS kernels" (they've existed since Feb 17)
• Added status table: Phase 1 DONE (perf gate FAIL), Phase 3 DONE (PR Unify GPU schedulers into a single coordinator thread #101), Phase 2/5 DEFERRED
• Updated existing assets table with GPU rANS kernels
• Added recommended next actions based on current evidence

Test plan

• No code changes — docs only
• Pre-commit hook passes (fmt, clippy, tests)

🤖 Generated with Claude Code