fix(badge): recognize path- and interpreter-prefixed okfmem sync as a save

[s-annam]

Summary

The save-state badge sat amber (okfmem*, "unsaved") over sessions that had saved and pushed — including the save that shipped the badge fix (#45) itself. SAVE_BASH_RE matched only the bare okfmem sync, but okfmem is an extensionless Python script, so the real close-out is python3 ~/okfmem/okfmem sync (and on Windows the .ps1/.cmd wrappers from #39). runs_command() anchors at a segment start after LEAD_RE strips env assignments and sudo/command/nohup — neither an interpreter wrapper nor a path prefix was stripped or allowed, so every non-bare shape of the save was invisible.

The direction is a false unsaved, benign per #44's framing, but systematic rather than incidental: it fired on every real save. An amber badge after a successful save is indistinguishable from one that means you forgot, which is exactly the signal the badge exists to carry.

Closes #46
Refs #44

What changed

• SAVE_BASH_RE allows a path prefix and the Windows wrapper extensions: (?:\S*[/\\])?okfmem(?:\.(?:ps1|cmd))?\s+sync\b.
• LEAD_RE additionally strips a leading interpreter and its flags (python3/py/pwsh/powershell/bash/sh/zsh, optionally path-qualified or .exe-suffixed), so python3 ~/okfmem/okfmem sync resolves to ~/okfmem/okfmem sync.

Command position is still required — the prefix is stripped, the anchor is not moved. Stripping an interpreter cannot manufacture a save, because what follows it must still be the program. Quoted spans are blanked before segmenting as before, so bash -c "okfmem sync" remains the documented accepted miss from #44.

An aggravating factor worth recording, not fixed here: /okfmem-save Writes its memory pages, MEMORY.md and STATE.md after its Skill{skill: okfmem-save} record, so the skill-invocation save signal is always older than the skill's own work. The terminal okfmem sync is the only signal that can legitimately flip the badge green — and it was the one being missed. The tie-to-unsaved policy stays as designed.

Measured

Session 8c03bb18 (the one that shipped #45), by transcript line index:

line	event	before	after
742	Skill{skill: okfmem-save}	save	save
788–877	Write memory page, Edit MEMORY.md, Write STATE.md, Write note	work	work
918	python3 ~/okfmem/okfmem sync -m "…"	missed	save

work_pos=877 > save_pos=742 → unsaved. After: saved.

Test plan

• python3 scripts/check-leaks.py exits 0 — Leak gate OK — scanned 62 tracked files, no leaks.
• python3 -m pytest tests/ -q — 199 passed, 1 skipped (2 new cases in tests/test_save_state_badge.py, one per direction)
• Save shapes all register: okfmem sync, ~/okfmem/okfmem sync, /abs/path/okfmem sync, ./okfmem sync, python3 ~/okfmem/okfmem sync -m "msg", okfmem.cmd sync, pwsh -File ./okfmem.ps1 sync, cd /r && python3 ~/okfmem/okfmem sync
• No false green: py build.py --then okfmem sync, grep -o 'okfmem sync' f, echo "run okfmem sync" still classify unsaved
• git commit variants from Save-state badge false 'saved': compute_save_state regexes raw transcript text, not structure #44 unregressed (git -C /repo commit --amend, cd /r && FOO=1 git commit, sudo git commit); grep -o 'git commit' log.txt still not work
• Replayed 8 real transcripts — sane spread (6 saved, 1 unsaved, 1 None), no blanket flip to green
• ruff check . — not run, ruff is not installed on this machine; CI covers it (advisory)

Provenance

Code implementation via: Claude Opus 4.8 (1M context), high effort
Verification: CI verify — pending