Reuse analysis across receipt worker batches

[chubes4]

Summary

• add compilePreparedPages() to compile independent page plans with one bounded worker-local analysis cache
• preserve compilePreparedPage() as the single-plan entrypoint
• report per-receipt transform counts rather than cumulative worker counts
• exercise interrupted 50-page resume through two worker batches
• verify batch receipts equal independently compiled receipts apart from elapsed-time metrics

Corpus evidence

On 58 persisted plans from the 203 MB corpus:

• eight isolated workers: approximately 1,500s wall and 10,970.45 aggregate worker-seconds
• four batched workers: 1,362.47s wall and 4,842.03 aggregate receipt-seconds
• aggregate CPU reduction: 55.9%
• slowest batch: 1,362.47s
• exact-version isolated/batched control receipt: byte-for-byte equal after removing compile_duration_ms

The API still returns one independently digest-bound receipt per page; batching changes worker-local analysis reuse, not terminal composition or persistence boundaries.

Addresses #1044.

Testing

• composer validate --strict
• composer test
• 280 parity fixtures
• WordPress and staged compilation contracts
• clean package-install proof
• 3-page exact receipt equality
• 50-page arbitrary-order resume and whole-result equality

AI assistance

OpenAI GPT-5.6 Sol via OpenCode assisted with corpus benchmarking, API design, implementation, and compatibility verification. Chris Huber reviewed and is responsible for the changes.