Doc lifecycle

Scope: YYYY-MM-DD-*.md files directly inside a project's notes dir (private_notes_ref.dir, default docs/_local, resolved the same way notes-sync.sh resolves it - git config --get notes-sync.dir, falling back to the literal default). A named state file with no date prefix (state-ledger.md, a ledger, a watermark file) is out of scope by construction, not a violation.

Terminal folder: <notes_dir>/archive/ (flat, no year subdirs). A doc's STATUS is terminal if and only if it lives in archive/; it is open if it lives directly in the notes dir.

doc-style's STATUS: line is documented there as optional; inside a notes dir it is mandatory (LIVE, SHIPPED, SUPERSEDED by <doc>, RECORD, written <date>, BLOCKED on <what> - same vocabulary, no new value). This is a narrowing for one location, not a contradiction between the two skills.

Creation: new dated docs are created with notes-sync.sh new <slug> (never by hand) - that command stamps today's real UTC date and a STATUS: LIVE line so both are correct by construction, closing the actual bug that produced 100% non-compliance in one repo's existing docs (a hand-typed date can misdate, and a hand-typed STATUS: line is easy to skip). --from <path> graduates an existing file's content into the notes dir the same way.

Close ritual - a checklist a human or a Lead follows manually, nothing here is scripted, every step is a judgment call:

1. Re-read the doc's open items.
2. Move them by content into whatever doc is current/live - never leave a pointer that requires re-opening the closed doc for active work.
3. Rewrite the closed doc's line 1 to a terminal STATUS, plus one line naming where open items went.
4. mv into archive/. This step always happens regardless of the disposition decision below - archive is the resting place either way, tracked or not.

Disposition (tracked vs stays private) is a SEPARATE decision, and it belongs to the user, never to a Lead acting on this checklist. Default: stays private, never tracked, even after closing - a doc having existed, or having been the plan that shipped, is never itself sufficient reason to track it. Skip this entirely unless the user raises it. When the user does decide something is worth keeping visible, promotion means copying the specific content into a real tracked doc (README, ADR, actual docs/) - it is never a byproduct of the close ritual above, and it never means leaving the original file tracked in place. A Lead must never run notes-sync.sh migrate, git rm --cached, or any other untrack/promote action on its own initiative off the back of a doc-lifecycle-check.sh finding - that report is evidence for the user's decision, never an instruction queue.

Advisory check, never a guard, in either direction: doc-lifecycle-check.sh never blocks a commit, a write, or anything else - every finding is a candidate for the user's own disposition decision. Three entry points:

• Standalone: scripts/doc-lifecycle-check.sh in any repo, any time - always exits 0, full list.
• SessionStart advisory (resume-inject.sh block E): passive, automatic, a capped summary only.
• /agent-tiers:doctor step 4e: the explicit "check this repo now" ask, full list shown inline.

Outstanding-work ledger (one per project, docs/_local/*outstanding-work-ledger*.md, line 1 STATUS: LIVE)

A deferral is a decision with a cost, so it is recorded where a reader can test it. One entry is one top-level - **title** bullet with these label lines (each at the start of its own line):

• Deferred: what is not being done. Marks the bullet as an entry.
• Cost: what it costs while undone, in countables where possible.
• Forced by: the concrete trigger, or the literal nothing named. Optionally one typed token the reader can evaluate: [session: S5] (fires when that plan row is DONE), [exists: path], [date: YYYY-MM-DD], [grep>=N: regex @ path]. Anchor it. "The next publish" is true the day it is written and unreadable a week later: nobody can tell whether that next has already passed. Name what it is next after (a date, a sha, a version) or use a typed token (a count goes in a [grep>=N: ...] token), and give the command that checks it. ledger-check flags an untyped trigger saying "next" with no anchor (L6).
• Probe: the cheapest check that would settle it now, and its result or not run.
• Source: where it was found (a review, a session, a path:line). Retire with a line-start DONE <date> [<session id or sha>]: <the check that proves it>. Never delete the line. A DONE with nothing after the colon, or carrying uncommitted, PARTIAL, except or pending, is not a retirement. Asked-for work is never deferred without a human: Narrowed: plus an AskUserQuestion at the time. A Deferred: line is for an out-of-scope find whose trigger cannot fire before the session ends; one that can fire is a fork, not a deferral.

scripts/ledger-check.sh (SessionStart, advisory) reads every live ledger: typed triggers that now hold, untyped trigger lines naming a DONE plan session, unanchored "next" triggers, DONE lines with no check or a qualifier, cited shas not in HEAD's history, DONE lines for a plan row that is not DONE, and the cap. Prose triggers are counted, never judged; entries with no labels are counted as legacy and listed only when their trigger line names a finished session. The plan it reads, when present: the first live docs/_local/*.md with a ## Status table heading over rows | <session id> | <state> | <proof> |, a row finished when its state starts DONE. Proof text (a DONE line, a status row's proof cell) names the command that proves it. A count it quotes is the result at close and says so (41/41 cases at close); a bare count reads as the current figure and goes stale the next time the suite grows. Never refresh an old count: the current figure is the command, re-run. Cap: 1500 lines or 150 KB per ledger. Past it, move closed entries older than 14 days to <name>-archive.md (STATUS: ARCHIVE), DONE lines verbatim; still over the cap means too many open entries, which is a human triage, not a formatting fix. Kill switch .state/off/ledger-check.