Comprehensive dedup quality: type/project/source awareness

[CalebisGross]

Summary

Systematic fix for false memory deduplication. The encoding agent was merging distinct memories that shared vocabulary (e.g., two different decisions about SQLite, or a decision vs an error about the same topic).

Changes

Rule	Before	After
Base threshold	0.9	0.95
MCP threshold	0.9 (same)	0.98 (higher for explicit input)
Cross-type	Allowed	Blocked (decision != error)
Cross-project	Allowed	Blocked (mnemonic != felix-lm)

Evidence

From the real-LLM lifecycle baseline (#262):

• 765 memories deduped at threshold 0.9
• 355 (46%) were between 0.9-0.95 — false merges
• Lowest dedup score: 0.9005 (barely over threshold)
• Average dedup score: 0.952

Architecture

Replaced bare float32 threshold with dedupContext struct:

type dedupContext struct {
    Threshold    float32 // base (default 0.95)
    MCPThreshold float32 // for explicit input (default 0.98)
    RawSource    string  // incoming memory source
    RawType      string  // incoming memory type
    RawProject   string  // incoming memory project
}

Both encoding paths (polling + event-driven) use buildDedupContext() → findDuplicate().

Test plan

• 10 unit tests (up from 4) covering all dedup scenarios
• make build + make test pass
• golangci-lint run — 0 issues
• Cross-type, cross-project, MCP threshold all tested

Closes #266