Pre-1.0 correctness and API hardening

[akshay-xp]

Pre-1.0 correctness and API hardening from the consolidated review. All on the stable v3 format; the only behavioral changes are validation and the blocked FPR mapping.

Changes

• C (BlockedBloomFilter.create: closed-form epsilon-to-bits/key mapping #99): BlockedBloomFilter.create sizes bits-per-key by solving the split-block false-positive rate in closed form instead of interpolating a 3-point anchor table. It now hits the target rate across the full epsilon range (eps=0.1 was silently ~18% before); targets below the solvable floor throw ParamError. Docs note the blocked-vs-classic crossover below ~1e-5.
• A (Validate constructor params against serialized field widths #100): constructors validate each parameter against the width it serializes into (k uint16, m/capacity uint32) via a new assertUint16, so out-of-range values throw ParamError instead of silently truncating.
• D (toBytes: single-allocation frame write path #101): toBytes writes into a single allocated frame through a new writeFrame primitive; the separate body buffer and copy are gone. Output is byte-identical (golden fixtures unchanged).
• E (Expose numBlocks/seed accessors on blocked and fuse #102): new numBlocks/seed accessors on BlockedBloomFilter and seed on BinaryFuse8/BinaryFuse16 for compatibility prechecks.
• F (Rename Hash128 fields to w0..w3 #103): internal Hash128 fields renamed h1lo/h1hi/h2lo/h2hi -> w0/w1/w2/w3 to match the x86_128 four-word layout.

Verification

• 162 tests passing; typecheck, lint, format, and size:check clean.
• Golden byte-identity preserved across all structures; hasher reference vectors unchanged.
• docs/api + API report regenerated; minor changeset added for the 0.6.0 batch.

Closes #99
Closes #100
Closes #101
Closes #102
Closes #103