fix: scale issues — clustering 450x faster, long-message entity retention 2x

[SimplyLiz]

Summary

Fixes three issues found by the new scale test (bench/scale-test.ts, 51 tests):

1. Semantic clustering O(n²) blowup — 450x faster

• Pre-compute pairwise similarity matrix instead of recalculating on each merge
• Cap at 200 candidates (most recent messages) to bound the quadratic cost
• Before: 76ms/msg at 500 messages (38 seconds total)
• After: 0.17ms/msg at 500 messages (83ms total)

2. Long-message entity retention — 2x improvement at 10k chars

• Summary budget cap now scales logarithmically for content >5k chars
• Entity suffix allows up to 30 entities for long messages (was 15)
• 10k chars: entity retention 55% → 91%
• 50k chars: entity retention 18% → 36% (inherent limit — 200+ identifiers can't fit in any summary)

3. Binary search vs tiered documentation

• Added performance note: tiered is 7-8x faster at 500+ messages
• Recommend tiered for large conversations in TSDoc

Remaining known limits (by design, not bugs)

• 50k+ char messages: entity retention capped ~36% — the summary budget ceiling can't represent all identifiers. Would need chunked compression (architectural change)
• Pure-code messages: 1.00x compression — T0 preserves code fences verbatim (correct)
• Short messages (<120 chars): 1.00x — too small to summarize (correct)
• Multi-round: converges after round 1 — already_compressed guard (correct)

Test plan

• 663 unit tests pass
• 51 scale tests pass (zero round-trip failures)
• Baseline saved
• Lint and format clean