Task (deferred): implement procedural memory mining (Procedure nodes)

[antejavor]

Part of #259

Question

Deferred by the driving user on 2026-08-11 -- do not implement until picked up again. The design is already decided in #262 (two separate mining processes -- sequence-mining over actions-graph, no LLM; content-mining over touched files, LLM-backed -- both writing a new Procedure node type, distinct from Skill, auto-write with no approval gate). What's still open before implementation can start:

• Sequence-mining's actual matching algorithm: window size, exact-vs-fuzzy tool-name sequence matching, and the repetition threshold before a sequence becomes a mined Procedure (a first attempt at this question was interrupted mid-clarification when this ticket was opened instead).
• Content-mining's file-touch tracking: actions-graph doesn't currently have an explicit "files touched during this session" abstraction -- would need to be inferred from tool_input on file-editing tool calls (Read/Write/Edit-style), which isn't designed yet.
• Where the code lives (new module in actions-graph, or a new dedicated package).

Resolve with the driving user via /grilling + /domain-modeling when picked back up.