fix(skill-store): fix broken embed calls and dedup threshold

[jfeng18]

Summary

The skill distillation storage pipeline (add_skill(), search_skills()) was non-functional due to two compounding bugs:

• self._embed() was never defined on Memory — all calls raised AttributeError at runtime. Replaced with self.embedding.embed() at 5 call sites.
• Dedup similarity_threshold defaulted to 0.75, but the skill store's RRF fusion score maxes out at 4*(1/60) = 0.0667. The dedup condition could never trigger, making the merge path dead code. Changed default to 0.03.

Details

Bug 1: _embed undefined

# Before (crashes):
title_emb = self._embed(title)        # AttributeError

# After (works):
title_emb = self.embedding.embed(title)  # uses the initialized embedder

Bug 2: threshold/score-space mismatch

Max RRF score: 4 * 1/(60+0) = 0.0667
Old threshold: 0.75  →  0.0667 > 0.75 = False (always)
New threshold: 0.03  →  0.0667 > 0.03 = True  (when 2+ lanes match)

Test plan

• Verified self.embedding is unconditionally set in __init__ (line 268)
• Verified embed() signature is compatible across all 16 embedder implementations
• Verified no remaining self._embed( calls in codebase
• Verified no test mocks _embed (no breakage)
• No existing tests assert on old threshold value 0.75
• CI tests pass
• Note: add_skill() and search_skills() have no dedicated test coverage (pre-existing gap)

🤖 Generated with Claude Code

[lightzt99]

LGTM.