feat: MaxEvidenceTokens and fixed-budget candidate pools

[sdntsng]

Summary

Wave D: LME subjects have ~240 extracts and untyped conversation_episode rows currently get +0.1 episode_boost, which promotes assistant congratulations/listicles over typed facts (gold “4 weeks” vs rank-1 “3 weeks”). MaxEvidenceTokens on /recall was dead.

This is PR4 (retrieval V4 at fixed context tokens). Kill-list compliant: no unbounded top-k; default overfetch cap stays 200.

Changes

• Wire MaxEvidenceTokens as the evidence assembly budget when set (else BudgetTokens / 4000).
• SearchOptions.CandidateLimit + RecallRequest.CandidateLimit; CandidatePoolSize qualifies 30/50/100/200 and caps at 200.
• Replace episode +0.1 boost with -0.15 penalty so typed facts outrank congratulation episodes.
• Tests: token truncation, candidate matrix, episode vs “4 weeks” fact, trace records pool size.

Test plan

• go test ./internal/memory/
• OpMem 13/13 + marketing after merge
• Do not treat this as a LoCoMo/LME quality lift until a dedicated pin on this binary

Linear / tracking

Wave D after #100. Independent of PR5 (#102) and PR3 (#103) on dev.

Risk / follow-ups

• Dated episodes still win date queries via date_token_boost / exact span (existing tests).
• Live 4-way LoCoMo candidate matrix is a follow-up pin, not this PR.
• No SOTA claim. dev only; do not merge to main.