[bug] procedural learning captures nothing: every session entry is written with an empty toolsUsed #591

What happened

Procedural learning has never produced a single skill. procedural_skills holds 0 rows in the maintainer's PostgreSQL store (checked 2026-09-16), and recall_skills returns {"skills": [], "count": 0} both cwd-scoped and unscoped. The same measurement was taken on 2026-08-25 during a harness comparison, so this is not a recent regression.

The chain stops at its first link. handlers/record_session_end.py:369 mines from load_session_log()["sessions"], and hooks/session_lifecycle.py:184-185 builds each of those entries as:

"turnCount": event.get("turn_count", 0),
"toolsUsed": event.get("tools_used") or [],

The SessionEnd payload carries session_id, transcript_path and cwd, not turn_count or tools_used. Measured on the live log (~/.claude/methodology/session-log.json): 413 sessions, every one with toolsUsed: [] and turnCount: 0. procedural_skill_writer._session_log_to_mining_input skips an entry with no tools, so the mining input is always empty, maybe_mine_skills returns skipped / insufficient_history, and nothing is ever written.

The transcript the event does carry has the evidence. Measured over 60 randomly sampled transcripts under ~/.claude/projects/ on 2026-09-16: 35 of 60 carry at least 2 tool uses (the MIN_SKILL_LEN floor), the median non-empty one carries 34, the largest carries 539, and extracting the sequence costs about 6 ms per transcript (0.4 s for all 60; a 17 MB transcript parses in 0.04 s).

A second defect sits one step further: infrastructure/scanner_parse.py:150 records "toolsUsed": list(stats["tools_used"]) where stats["tools_used"] is a set, so the import path produces an unordered, deduplicated list. core/procedural_memory.mine_skills mines contiguous subsequences, so an unordered list yields sequences that were never performed.

Expected behavior

A session entry carries the ordered tool sequence actually performed, with repetitions, derived from the transcript when the event does not supply it, and a turn count. Mining then has real input, and recall_skills can return something.

Evidence required to close

A test deriving the sequence from a transcript fixture (order and repetition preserved), a test that the session entry is filled when the event omits the fields, and an end-to-end measurement showing mine_skills producing at least one skill from real session entries.