perf: distillation loses specific identifiers at 400K+ token sessions

[BYK]

Problem

At 400K tokens, Lore scores 3.5 vs compaction's 4.2 on CM-1. The gap is specifically on medium-difficulty questions where distillation dropped specific details (exact error messages, alternative approaches considered, config file paths).

Root Cause

Lore's two-stage compression pipeline loses more information than compaction's single-pass summary:

1. Gen-0 compression: Each 16K segment gets 8*sqrt(16384) = 1024 tokens. This is 16:1 compression — specific identifiers (error messages, file paths, version numbers, rejected alternatives) are dropped.
2. Meta-distillation: 20 gen-0 segments (~20K tokens) compressed to ~5-6K tokens. Another 3-4:1 compression compounding the loss.
3. Total: 320K → 20K → 5-6K = ~53-64:1 compression vs compaction's 320K → 4K single-pass (80:1 but higher quality due to seeing raw content).

Evidence

CM-1 inflated to 400K tokens:

Baseline	Easy	Medium	Hard	Overall
Lore	5.0	2.3	3.3	3.5
Tail-window	4.7	1.3	1.4	2.5
Compaction	4.7	3.9	4.1	4.2

Failed medium questions: model says 'no alternative approach was explicitly discussed' (it was), 'I don't have information about an integration test failure' (there was one).

Possible Fixes

1. Increase gen-0 budget multiplier from 8 to 10-12 (more tokens per segment)
2. Raise the CAP from 4096 to 6144 or 8192 (allows larger segments more budget)
3. Skip meta-distillation for recent segments (keep gen-0 detail for the most recent N segments)
4. Improve observer prompt to prioritize specific identifiers over chronological event logging
5. Increase tool output truncation limit from 2K chars (distillation input) — specific error messages in tool results get truncated before the LLM sees them

Context

Discovered during #410 investigation. The eval fix (#414) correctly represents Lore's behavior — this gap is a real Lore limitation, not an eval bug.