feat: infer tags from recall query for soft ranking boost (#142)

[rahilp]

Summary

• Adds inferQueryTags(query, env): three-tier tag inference — hashtag extraction (free, no DB), keyword match against known tag vocabulary (json_each query), and LLM fallback only when both find nothing; runs in parallel with embedding via Promise.all so the common path adds no latency
• Adds queryTags: string[] param to rerankWithTimeDecay: entries whose tags overlap with inferred tags get a soft boost (TAG_BOOST_STEP=0.15, TAG_BOOST_MAX=1.5) applied outside the recency cap — tag-relevant memories can surface above marginally-closer but irrelevant ones without over-narrowing results
• Wires both into recallEntries: explicit tag filter behaviour is unchanged; inference is soft ranking only

Resolves #142

Test plan

• npm test — 314/314 passing
• npx tsc --noEmit — no TypeScript errors
• Manual: recall "what did I decide about #work today" — work-tagged memories rank higher
• Manual: recall "office lease" with entries tagged "legal" — legal-tagged memories rank higher without explicit tag filter
• Manual: recall with tag=work filter — no regression in existing tag-scoped behaviour

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	82.73%	604 / 730
🔵	Statements	83.6%	714 / 854
🔵	Functions	82.75%	96 / 116
🔵	Branches	77.29%	412 / 533

File Coverage

File	Stmts	Branches	Functions	Lines	Uncovered Lines
Changed Files
src/index.ts	82.56%	75.41%	81.48%	81.73%	113-160, 180, 471, 478, 519, 784, 895, 901, 910, 918-938, 1022-1025, 1057, 1064, 1065, 1076, 1098-1106, 1116, 1168, 1209, 1222, 1268-1490, 1499-1503, 1516-1541, 1560, 1602, 1640, 1674, 1840-1851, 1913, 1923

Generated in workflow #141 for commit 76a206d by the Vitest Coverage Report Action