fix: improve LTM preference injection — skip relevance scoring, dedicated budget, meaningful confidence

[BYK]

Summary

• Preference-only fast path in forSession(): skips FTS5/vector relevance scoring for preferences (which are unconditional user directives, not contextual knowledge). Sorts by confidence DESC, updated_at DESC with no count cap — fixes cross-project preferences being silently dropped.
• Dedicated 2% preference budget independent of the 5% context-bound LTM budget, so preferences no longer compete with gotchas/patterns for token space.
• Meaningful confidence on create: ltm.create() and CuratorOp create ops now accept confidence. Curator prompt updated with explicit semantics (1.0=unconditional directive, 0.9=strong preference, 0.8=moderate, 0.6=mild).
• lore data rerank CLI command to re-score legacy preference entries by directive-detection patterns. Also auto-runs after lore data recover.

Problem

Most user preferences were not injected into sessions because:

1. Cross-project preferences had no safety net in forSession() and were dropped when they scored 0 on relevance (e.g., "never push to main" has zero FTS5 overlap with a React session)
2. NO_CONTEXT_FALLBACK_CAP=10 capped first-turn preferences
3. Preferences competed with context-bound entries for the same 5% budget
4. confidence was always hardcoded to 1.0, making it useless for ranking

Files Changed

File	Change
packages/core/src/ltm.ts	Preference fast path, confidence param on create(), rerankPreferences()
packages/core/src/curator.ts	confidence on CuratorOp create type, wired through applyOps
packages/core/src/prompt.ts	Curator prompt with confidence semantics
packages/core/src/config.ts	New preferenceLtm budget field (default 0.02)
packages/core/src/gradient.ts	getPreferenceLtmBudget() function
packages/core/src/index.ts	Export new function
packages/gateway/src/pipeline.ts	Dedicated preference budget, simplified context-bound budget
packages/gateway/src/cli/data.ts	lore data rerank command, auto-rerank after recover
packages/core/test/ltm.test.ts	11 new tests covering fast path, confidence, and reranking

Testing

• 1553 tests pass, 0 failures
• All 4 packages typecheck clean
• After deploying, run lore data rerank to fix existing preference entries