knowledge: Co-occurrence clustering for project inference

[timkicker]

Goal

Implement the fourth membership mechanism described in docs/architecture/project-system.md and foundation/sections/04-event-knowledge.tex §4.7: detect coherent project clusters from co-occurring file accesses, terminal working directories, and (eventually) browser tabs that consistently appear together in sessions.

Background

The signal-based detection in knowledge/src/project/signals.rs covers the first three mechanisms (explicit .project, rootdir heuristic, manual assignment). Co-occurrence clustering is unimplemented (0 grep hits across knowledge/).

The blueprint frames this as: "files that consistently appear together in the same sessions, accessed by the same applications and terminal working directories, are grouped as project candidates by a background inference pass."

Scope

• Background inference task (separate from the existing projects task — runs at lower frequency, e.g. daily)
• Reads recent Session + ACCESSED_BY + ACTIVE_IN edges from the graph
• Identifies clusters via a co-occurrence frequency metric (specific algorithm TBD: simple threshold, Jaccard similarity, or affinity propagation)
• Creates Project nodes with inferred = true, promoted = false, confidence set from cluster strength
• Surfaces candidates as "suggestions" — does NOT auto-link files via PART_OF
• User confirms via Waypointer to promote (see also #95 Save-as-Project action)

Out of scope

• Browser tab integration — needs browser extension, defer to post-Phase 8
• ML-based clustering — start with deterministic frequency heuristic

References

• docs/architecture/project-system.md §File membership and the PART_OF edge (mechanism Implement Graph Daemon Cypher query interface #3)
• foundation/sections/04-event-knowledge.tex §Project System

Acceptance criteria

• New co_occurrence module under knowledge/src/project/
• Inference task runs on schedule and produces candidate Project nodes
• Each candidate has confidence set (0-100) and is marked inferred + !promoted
• Tests cover: empty graph, single session no clusters, multi-session strong cluster, weak cluster below threshold

Phase

Phase 5 (Foundation cleanup follow-up). Discovered during 5A issue-hygiene pass — this was scoped under #11/#13 but never built, so it gets its own ticket rather than reopening those.