Profile and bound lineage replay performance

[iam-brain]

Question

Reduce the lineage ledger runtime and memory cost enough for safe flagged promotion without weakening correctness. The frozen 819,779-observation replay currently takes materially longer than legacy scanning and peaks on a 355,168-observation family.

Acceptance

• Add stage-level timing and family-size diagnostics without private identifiers.
• Profile parsing, family preparation, fingerprinting, reconciliation, pricing, and fallback separately.
• Preserve family-local release and cancellation.
• Add bounded performance tests for multi-million-observation synthetic histories.
• Demonstrate a material frozen-corpus improvement before Remove superseded fork accounting and harden migrations #20; do not promote solely because unit benchmarks pass.

Prerequisite for #20 and successor to #23.

[iam-brain]

Implemented in stacked PR #44. The frozen 570-document replay preserved exact daily UTC totals and reduced lineage runtime from 224.9s to 163.8s in the instrumented comparison (about 54% faster than the earlier ~360s run). Family reconciliation fell from 72.4s to 15.2s. Discovery and document loading remain about 74s each and are the next optimization frontier. Full suite: 649 selections, 55/55 groups, no failures/retries/timeouts.