[P2][OCR] Reduce redundant context serialization and oversized string allocation

[Ryder-Sun]

Source: OCR_ANALYSIS_REPORT.md, OpenCode Review report generated on 2026-08-12.

Findings: M-5, L-2. The context-pack code serializes near-limit records more than once and allocates full UTF-8 buffers before rejecting oversized strings. This is not currently a correctness failure, but it increases CPU and memory pressure around the 2 MB context-pack boundary.

Expected fix:

• Reuse the canonical JSON string returned by size enforcement.
• Add a cheap pre-check before allocating full buffers for clearly oversized strings.
• Add performance-oriented regression coverage around near-limit payloads.