fix: split namespace on KV layout factors, bound GPU copies to stored segments

[xiaguan]

Context

Follow-up to the discussion in #344. Two of the three holes identified there are closed here; the engine-side seal-time fingerprint remains open for discussion in the issue.

Changes

1. derive_namespace: cover block_size and KV-bearing speculative drafters

The namespace hash missed two factors that change stored-block compatibility:

• block_size — deployments differing only in tokens-per-block derived the same namespace. In practice vLLM's block-hash grouping kept them from ever matching, but that is protection by accident, not by construction.
• Speculative drafters with their own KV cache (MTP, EAGLE, draft model) — the drafter adds registered KV layers, changing the block slot layout. MTP-on and MTP-off deployments of the same model shared a namespace: the warm side keeps reporting hits the other side can never serve (loud load failures since fix(core)!: seal instance layer topology from registered layers #346, perpetual recompute + error spam).

N-gram-style methods (ngram, ngram_gpu, suffix) have no drafter KV and keep sharing the namespace. Unknown methods conservatively split — a needless split is a cold cache, a missed split is cross-contamination.

Upgrade note: all existing namespaces change. This is a one-time cold cache: the SSD index is in-memory, metaserver entries expire via TTL, resident blocks age out through LRU. A golden-value test now pins the hash so future factor changes are explicit decisions instead of refactoring side effects.

2. build_copy_descs: validate stored block segments before CUDA copy

Copy sizes come from the instance's GPU layout, but the host block may come from a foreign source (shared namespace, SSD, RDMA remote) with different geometry. Previously there was zero size validation — an undersized stored segment meant an out-of-bounds read/write on the pinned pool, silent because the pool mapping is large.

Now every segment is checked against the bytes the layout will copy, and a contiguous-layout copy requires a single-segment host block (a multi-segment block stores V in segment 1, which one contiguous copy cannot serve even when segment 0's padding passes the size check). Legitimate save-path blocks always pass: host segments are allocated at padded_segment_bytes >= layout segment bytes.

Testing

• cd python && uv run --extra test pytest — 174 passed (new test_namespace.py covers MTP/EAGLE/unknown split, ngram share, block_size split, golden pin)
• Source-only gate (uv run --isolated --no-project --with pytest --with numpy --with 'requests>=2.26.0' pytest) — 174 passed
• cargo test -p pegaflow-core --no-default-features --features cuda-13 — all green
• clippy/fmt clean (pre-existing warnings unchanged)

Refs #344

🤖 Generated with Claude Code