dorfl 0.13.2 → 0.13.4

package/src/needs-attention.ts

@@ -19,6 +19,7 @@ import {
 	readItemLock,
 	itemLockRef,
 	lockEntryFor,
+	refreshMainRef,
 	parseLockEntry,
 	type LockEntry,
 } from './item-lock.js';
@@ -33,6 +34,7 @@ import {
 	resolveSidecarIdentity,
 	serialiseSidecar,
 	sidecarPathFor,
+	sidecarPathCandidates,
 	type NewQuestion,
 	type SidecarType,
 } from './sidecar.js';
@@ -2140,6 +2142,787 @@ export function prepareTreelessSurfaceCommit(params: {
 	}
 }
 
+/**
+ * The kind of TERMINAL resting place an item has reached on `main`, which
+ * decides how much of its question state is residue.
+ *   - `completed`, the work HAPPENED (`tasks/done/`). Any surviving question
+ *     state is pure residue: the questions were about how to proceed, and the
+ *     item proceeded.
+ *   - `wont-proceed`, the item was ABANDONED (`tasks/cancelled/`,
+ *     `specs/dropped/`). Here `needsAnswers:true` may be ACCURATE HISTORY: an
+ *     item can be cancelled precisely BECAUSE its questions were never answered,
+ *     and the body may carry a real `## Open questions` section recording that.
+ *
+ * NOTE what is ABSENT: `specs/tasked/`. It is a terminal RESIDENCE, but this map
+ * is keyed to "is the question loop CLOSED here?", not "has the item stopped
+ * moving?", and on a tasked spec the loop is explicitly still open (see the
+ * `case 'spec'` comment below).
+ */
+export type TerminalKind = 'completed' | 'wont-proceed';
+
+/**
+ * The terminal `work/` paths for an item, tagged by {@link TerminalKind}, so a
+ * reader can tell "the work happened" from "the item was abandoned".
+ *
+ * Same SHAPE as `terminalMainPaths` in `item-lock.ts`, but deliberately NOT the
+ * same folder set, and the difference must not be "tidied" away: locks treat
+ * `specs/tasked/` as terminal (correctly, a tasked spec must release its lock),
+ * whereas QUESTION state there is still live. This map is keyed to "is the
+ * question loop CLOSED at this resting place?", not "has the item stopped
+ * moving?". See the `case 'spec'` comment below.
+ */
+export function terminalMainPathsByKind(
+	type: SidecarType,
+	slug: string,
+): {path: string; kind: TerminalKind}[] {
+	const file = `${slug}.md`;
+	switch (type) {
+		case 'task':
+			return [
+				{path: workItemRel('done', file), kind: 'completed'},
+				{path: workItemRel('cancelled', file), kind: 'wont-proceed'},
+			];
+		case 'spec':
+			// `specs/tasked/` is deliberately NOT listed. WORK-CONTRACT ("A SPEC that
+			// has drifted AFTER it was TASKED") makes a bare `needsAnswers:true` on a
+			// tasked spec LEGAL and load-bearing: it means "tasked, but the spec has
+			// drifted, do not RE-task or rely on it until reconciled", and the
+			// contract says to set it *while the spec stays in `specs/tasked/`*
+			// (moving it back would falsely un-record a tasking that really happened
+			// and orphan the tasks it already emitted).
+			//
+			// So the reasoning that makes a task's question state moot at its terminal
+			// does NOT transfer: a tasked spec is still IN the question loop.
+			// `lifecycle-gather.ts` enumerates tasked resting specs UNCONDITIONALLY,
+			// routing a bare flag to the SURFACE rung and an answered sidecar to the
+			// APPLY rung, so BOTH halves are live inputs to a rung that WILL run.
+			// Draining either would disarm a live drift gate and let a stale spec be
+			// re-tasked. That is precisely the "clearing a live needsAnswers hands gated work
+			// to agents" harm this pass exists to avoid.
+			//
+			// `specs/dropped/` needs no such carve-out: a dropped spec is abandoned,
+			// and no rung enumerates it.
+			return [{path: workItemRel('specs-dropped', file), kind: 'wont-proceed'}];
+		case 'observation':
+			// A note has no durable terminal folder: it leaves by DELETION, so there
+			// is no resting record to reconcile against.
+			return [];
+	}
+}
+
+/** Every SUCCESS-terminal folder that can hold a stranded `needsAnswers` flag,
+ * paired with the item type that rests there. Derived from
+ * {@link terminalMainPathsByKind} with a sentinel slug so the folder set stays
+ * SINGLE-SOURCED: adding a regime there adds it here, and the `wont-proceed`
+ * terminals are excluded by the SAME `kind` split the drain already branches on
+ * (a cancelled item's flag is accurate history, not residue). `observation`
+ * contributes nothing, having no durable terminal. */
+function successTerminalFolders(): {folder: string; type: SidecarType}[] {
+	const out: {folder: string; type: SidecarType}[] = [];
+	for (const type of ['task', 'spec', 'observation'] as const) {
+		for (const candidate of terminalMainPathsByKind(type, '__slug__')) {
+			if (candidate.kind !== 'completed') {
+				continue;
+			}
+			out.push({
+				folder: candidate.path.slice(0, candidate.path.lastIndexOf('/')),
+				type,
+			});
+		}
+	}
+	return out;
+}
+
+/**
+ * A SUCCESS-terminal item carrying a STRANDED `needsAnswers:true` flag with NO
+ * sidecar beside it: the residue's harmful half, on its own.
+ *
+ * This is NOT the mirror state the classifier calls legal. `needsAnswers:true`
+ * with no sidecar IS normal on a POOL or STAGING item (it is precisely the
+ * `surface` rung's input, and clearing it there would disarm every un-surfaced
+ * gated item in the repo). What makes THIS shape residue is the POSITION: the
+ * item has already SHIPPED, so there is no question left to surface and no
+ * answer that could still be typed, because `surface` will never run on it again.
+ *
+ * It is reached whenever the two halves are separated in the one order the
+ * sidecar-anchored sweep cannot follow: the SIDECAR goes first and the FLAG is
+ * left behind. A human tidying `work/questions/` by hand does exactly that (the
+ * obvious manual clean-up, and the sidecar is the visible half), which is how
+ * the fix
+ * for the paired residue can report success while any gate it cannot see stays
+ * armed. Anchoring only on the sidecar set makes hand-cleanup permanently strand
+ * the half that actually gates work.
+ */
+export interface TerminalFlagResidue {
+	/** The namespaced identity (`task:<slug>`). */
+	item: string;
+	/** The item body's SUCCESS-terminal path on `main`. */
+	itemPath: string;
+}
+
+/** One item whose question state survived into a terminal resting place. */
+export interface TerminalQuestionResidue {
+	/** The namespaced identity (`task:<slug>`). */
+	item: string;
+	/** The sidecar's path on `main` (`work/questions/<type>-<slug>.md`). */
+	sidecarPath: string;
+	/** The item body's terminal path on `main`. */
+	itemPath: string;
+	/** Which terminal the body rests in. */
+	terminal: TerminalKind;
+	/** Does the body still carry `needsAnswers: true`? */
+	flagged: boolean;
+	/**
+	 * Does the sidecar carry at least one ANSWERED entry? Such a sidecar holds
+	 * human-written prose that was never consumed by the apply rung, so the drain
+	 * refuses to touch it (see {@link classifyTerminalQuestionResidue}).
+	 */
+	answered: boolean;
+}
+
+/** The read-only classification of the arbiter's stranded question state. */
+export interface TerminalQuestionReport {
+	/** Residue the drain WILL clear: terminal + no answered entry. */
+	drainable: TerminalQuestionResidue[];
+	/**
+	 * Residue the drain deliberately LEAVES: a terminal item whose sidecar carries
+	 * a human's ANSWER that was never applied. Reported for a human, never
+	 * silently deleted (the answer is data the tool did not author).
+	 */
+	answeredHeld: TerminalQuestionResidue[];
+	/**
+	 * SUCCESS-terminal items whose `needsAnswers` gate is armed with NO sidecar
+	 * beside it. Cleared by the drain (there is no sidecar, so nothing a human
+	 * wrote can be discarded). See {@link TerminalFlagResidue}.
+	 */
+	staleFlags: TerminalFlagResidue[];
+	errors: {item: string; message: string}[];
+}
+
+/**
+ * Classify the arbiter's STRANDED QUESTION STATE, read-only (observation
+ * `a-rebuilt-task-leaves-its-bounce-question-asking-to-cancel-a-merged-task`).
+ *
+ * THE BUG THIS EXISTS FOR. When a build bounces, the surface path atomically
+ * writes BOTH halves of the question state in ONE commit: the sidecar
+ * `work/questions/<type>-<slug>.md` AND `needsAnswers: true` on the item body.
+ * That is correct, and the atomicity is what makes this reconciliation decidable
+ * at all. But if the human DISAGREES with the bounce and simply re-dispatches,
+ * and the rebuild SUCCEEDS (PR opened, gate green, merged, body done-moved),
+ * NEITHER half is ever cleared. The item comes to rest in `tasks/done/` still
+ * carrying a question asking whether to CANCEL it, with a destructive default.
+ *
+ * The flag is the harmful half. A stranded sidecar is a stale question in a
+ * folder a human scans; a stranded `needsAnswers` is a GATE LEFT ARMED, and it
+ * makes `status` report shipped (sometimes released) work under "open questions
+ * block autonomous work".
+ *
+ * Dorfl ALREADY knows this state is illegal: `advance-classify.ts` refuses it as
+ * `invariant-violation` / `sidecar-without-needsAnswers`. The defect is purely
+ * that the detector lives in the `advance` tick's classifier, and a human driving
+ * `do` and merging a PR never enters that loop. So this is the same shape as the
+ * propose-path lock leak, settled by the same reconcile pass at the same moment
+ * (the done-move), rather than by a second mechanism.
+ *
+ * THE TRAP, and why the TERMINAL POSITION is the discriminator rather than the
+ * flag/sidecar disagreement: the MIRROR state (`needsAnswers:true` with NO
+ * sidecar) is LEGAL and COMMON. An item authored with open questions carries the
+ * flag and has no sidecar until `surface` runs, and that flagged-but-unsurfaced
+ * item is precisely the `surface` rung's INPUT. Clearing the flag there would
+ * silently disarm every un-surfaced item in the repo and hand gated work to
+ * agents. So this only ever considers items whose body has reached a TERMINAL
+ * folder on `main`; an item resting in a pool or staging folder keeps whatever
+ * state it has, untouched.
+ *
+ * TWO ENUMERATIONS, because the two halves can be separated in either order and
+ * a sweep anchored on one is blind to the other:
+ *   1. the SIDECAR SET (`work/questions/` on `main`), small and cheap to list,
+ *      which finds a stale sidecar and the flag paired with it; and
+ *   2. the SUCCESS-TERMINAL BODIES that are flagged with NO sidecar beside them
+ *      ({@link collectStrandedTerminalFlags}), which finds the armed gate ALONE.
+ *
+ * (2) is not optional tidiness. It is the half that actually gates work, and a
+ * sweep anchored only on (1) reports success while any gate it cannot see stays
+ * armed. The sidecar is the
+ * half a human deletes by hand (it is the visible one, in a folder they scan),
+ * and deleting it REMOVES the only handle (1) has, stranding the flag for good.
+ * The discriminator that keeps (2) safe is POSITION, exactly as for (1): a bare
+ * flag is LEGAL on a pool/staging item (the `surface` rung's input) and residue
+ * only once the item has shipped, where `surface` can never run again.
+ *
+ * Best-effort and never throws.
+ */
+export async function classifyTerminalQuestionResidue(params: {
+	cwd: string;
+	arbiter: string;
+	/** The ref holding the arbiter's authoritative `main`. */
+	mainRef: string;
+	env?: NodeJS.ProcessEnv;
+	/** Skip the `mainRef` refresh because the CALLER just did it (the combined
+	 * pass refreshes once and runs both sub-passes against that ONE snapshot). */
+	mainAlreadyFresh?: boolean;
+}): Promise<TerminalQuestionReport> {
+	const {cwd, mainRef, env} = params;
+	// REFRESH `mainRef` FIRST, with an explicit refspec that writes exactly the ref
+	// we are about to read. Without this the pass reads a STALE view: the caller
+	// may not have fetched, and the lock sub-pass of the combined reconciliation
+	// early-returns (so does not refresh) when no locks are held. A failed refresh
+	// is NOT fatal, but it does mean the view may be stale in EITHER direction (an
+	// item may have left a terminal folder, or acquired an answer, since we last
+	// looked), which is exactly why the WRITE path re-derives this same
+	// classification against its own freshly-resolved base rather than trusting
+	// this snapshot.
+	await refreshMainRef(mainRef, params.arbiter, cwd, env);
+	return deriveTerminalQuestionResidue(mainRef, cwd, env);
+}
+
+/**
+ * The SYNC, PURE-of-network derivation of the question residue AT ONE COMMIT.
+ *
+ * Split out of {@link classifyTerminalQuestionResidue} so the WRITE path can
+ * re-derive the SAME classification against the base it is actually about to
+ * commit on, per contention attempt. That matters for correctness, not tidiness:
+ * a classification taken before a contention retry can be stale in two ways that
+ * both break a documented guarantee. An item may have LEFT its terminal folder
+ * (re-opened), in which case its sidecar is live again and must not be deleted;
+ * and a human may have written an ANSWER into a sidecar in the window, which must
+ * never be auto-deleted. Re-deriving against `base` closes both, because the
+ * commit is built on exactly that base.
+ */
+function deriveTerminalQuestionResidue(
+	base: string,
+	cwd: string,
+	env: NodeJS.ProcessEnv | undefined,
+): TerminalQuestionReport {
+	const mainRef = base;
+	const out: TerminalQuestionReport = {
+		drainable: [],
+		answeredHeld: [],
+		staleFlags: [],
+		errors: [],
+	};
+	// The SECOND half of the residue, enumerated from the OTHER side. The sidecar
+	// sweep below can only ever see items that still HAVE a sidecar; this one finds
+	// the SUCCESS-terminal bodies whose gate is armed with no sidecar left to point
+	// at them. Both must run: they are the same defect observed through the two
+	// halves the surface path writes atomically, and either half can outlive the
+	// other.
+	collectStrandedTerminalFlags(mainRef, cwd, env, out);
+	const questionsDir = workFolderRel('questions');
+	const ls = run(
+		'git',
+		['ls-tree', '--name-only', `${mainRef}:${questionsDir}`],
+		cwd,
+		{env},
+	);
+	if (ls.status !== 0) {
+		// No `work/questions/` on main at all: nothing surfaced, nothing to drain.
+		return out;
+	}
+	for (const name of ls.stdout.split('\n').map((l) => l.trim())) {
+		if (name === '' || !isWorkItemFile(name)) {
+			continue;
+		}
+		const sidecarPath = `${questionsDir}/${name}`;
+		try {
+			// `<type>-<slug>.md` → `<type>:<slug>`. Only the CURRENT namespaces are
+			// addressable; a legacy `prd-` file has no current item-form and is left
+			// for the migration command.
+			const stem = name.replace(/\.md$/, '');
+			const dash = stem.indexOf('-');
+			const type = stem.slice(0, dash) as SidecarType;
+			const slug = stem.slice(dash + 1);
+			if (!['task', 'spec', 'observation'].includes(type) || slug === '') {
+				continue;
+			}
+			const item = `${type}:${slug}`;
+			// Is the body at rest in a terminal folder on `main`?
+			const terminalHit = terminalMainPathsByKind(type, slug).find((c) =>
+				pathInCommit(mainRef, c.path, cwd, env),
+			);
+			if (terminalHit === undefined) {
+				// NOT terminal: a live item. Its question state is its own business
+				// a pending sidecar is a human's outstanding decision, and clearing a
+				// flag here is the trap above. Untouched.
+				continue;
+			}
+			// N5 GUARD: a mid-migration spec can have BOTH `spec-<slug>.md` and the
+			// legacy `prd-<slug>.md` on main (`sidecarPathCandidates` still resolves
+			// the legacy name for readers). Draining only the canonical one while
+			// clearing the flag would leave the legacy sidecar live against
+			// `needsAnswers:false`, which is precisely the
+			// `sidecar-without-needsAnswers` invariant violation this change exists
+			// to remove. If any OTHER candidate for this item still exists, leave the
+			// whole item to `dorfl prd-to-spec`, which renames the DATA.
+			const hasLegacyAlias = sidecarPathCandidates(item).some(
+				(c) => c !== sidecarPath && pathInCommit(mainRef, c, cwd, env),
+			);
+			if (hasLegacyAlias) {
+				continue;
+			}
+			const model = parseSidecar(
+				catBlob(`${mainRef}:${sidecarPath}`, cwd, env),
+			);
+			const answered = model.entries.some((e) => isEntryAnswered(e));
+			const body = catBlob(`${mainRef}:${terminalHit.path}`, cwd, env);
+			const flagged = parseFrontmatter(body).needsAnswers === true;
+			const residue: TerminalQuestionResidue = {
+				item,
+				sidecarPath,
+				itemPath: terminalHit.path,
+				terminal: terminalHit.kind,
+				flagged,
+				answered,
+			};
+			if (answered) {
+				// A human WROTE an answer here and the apply rung never consumed it.
+				// Deleting it would discard prose the tool did not author, so this is
+				// surfaced for a human instead. (That the drain never runs on the
+				// human-answer path either is a SEPARATE defect; this pass must not
+				// paper over it by destroying the evidence.)
+				out.answeredHeld.push(residue);
+			} else {
+				out.drainable.push(residue);
+			}
+		} catch (err) {
+			out.errors.push({
+				item: sidecarPath,
+				message: err instanceof Error ? err.message : String(err),
+			});
+		}
+	}
+	return out;
+}
+
+/**
+ * Find every SUCCESS-terminal body on `base` carrying `needsAnswers:true` with NO
+ * sidecar beside it, appending them to `out.staleFlags`.
+ *
+ * ENUMERATION COST is why this is a `git grep` and not a walk. The terminal
+ * folders are the repo's largest and most monotonically growing (this repo holds
+ * 404 done tasks), and this runs on the CLAIM path, so reading every terminal
+ * body per claim would be a real tax on a hot path. One `git grep -l` returns
+ * only the candidates, and the frontmatter parse runs over that short list.
+ *
+ * The pattern is ANCHORED to match the PARSER rather than the word.
+ * `parseFrontmatter` reads keys with `/^([A-Za-z0-9_.]+)\s*:\s*(.*)$/`, so a key
+ * it will honour is always at column 0; an unanchored needle instead matches
+ * every body that merely DISCUSSES the flag, which in `work/tasks/done/` here is
+ * 77 files against 18 anchored, and the truthy form narrows it to 1.
+ *
+ * The value part is matched LOOSELY on purpose (optional quote, any case),
+ * because `toBoolean` unquotes and lower-cases before comparing, so
+ * `needsAnswers: 'True'` is a real armed gate. A needle of `:\s*true` would read
+ * tighter and be WRONG: it would silently skip those bodies for ever, which is
+ * the blind-spot class this function exists to remove. A superset is the safe
+ * direction for a shortlist; a subset is not.
+ *
+ * The grep is still only a CANDIDATE FILTER, never the decision: prose can sit
+ * at column 0 too (`needsAnswers: true?` appears in this repo's own bodies), so
+ * every hit is confirmed by actually PARSING the frontmatter.
+ *
+ * Two git-isms are pinned rather than left to the environment:
+ *   - `core.quotePath=false`, or git C-quotes any non-ASCII path
+ *     (`"work/.../caf\303\251.md"`). A quoted line still starts with the
+ *     `<base>:` prefix but then fails the folder-prefix test, so such a body
+ *     would be SILENTLY skipped for ever, a permanent blind spot of exactly the
+ *     class this function exists to remove.
+ *   - `--full-name` + `:(top,literal)` pathspecs, because `git grep`'s pathspecs
+ *     are CWD-RELATIVE (unlike the `ls-tree`/`cat-file` probes elsewhere here,
+ *     which are tree-relative) and are globs. Without these, running any dorfl
+ *     command from a SUBDIRECTORY makes this half a silent no-op while the
+ *     sidecar half keeps working.
+ *
+ * Never throws; a failed grep yields no candidates, which leaves state alone.
+ */
+function collectStrandedTerminalFlags(
+	base: string,
+	cwd: string,
+	env: NodeJS.ProcessEnv | undefined,
+	out: TerminalQuestionReport,
+): void {
+	const folders = successTerminalFolders();
+	if (folders.length === 0) {
+		return;
+	}
+	// `-l` names files only, `-I` skips binaries. Exit 1 means NO MATCH, which
+	// ALSO covers "the folder does not exist on this base yet" (verified: an
+	// absent pathspec folder exits 1 with no stderr), and an absent lifecycle
+	// folder is legal per WORK-CONTRACT rule 3. Any OTHER non-zero is a genuine
+	// fault and is REPORTED rather than swallowed: degrading silently to "no
+	// candidates" would leave this half a no-op while the sidecar half keeps
+	// reporting success, which is the very "reports success while the gate stays
+	// armed" shape this change exists to correct.
+	const grep = run(
+		'git',
+		[
+			'-c',
+			'core.quotePath=false',
+			'grep',
+			'-l',
+			'-I',
+			'--full-name',
+			'-E',
+			'^needsAnswers:[[:space:]]*[\'"]?[Tt][Rr][Uu][Ee]',
+			base,
+			'--',
+			...folders.map((f) => `:(top,literal)${f.folder}`),
+		],
+		cwd,
+		{env},
+	);
+	if (grep.status !== 0) {
+		if (grep.status !== 1) {
+			out.errors.push({
+				item: '(stranded-flag scan)',
+				message:
+					`git grep over the terminal folders failed (exit ${grep.status}): ` +
+					`${grep.stderr.trim() || 'no stderr'}; stranded gates were NOT scanned.`,
+			});
+		}
+		return;
+	}
+	const prefix = `${base}:`;
+	for (const line of grep.stdout.split('\n')) {
+		const raw = line.trim();
+		if (raw === '' || !raw.startsWith(prefix)) {
+			continue;
+		}
+		const path = raw.slice(prefix.length);
+		try {
+			const home = folders.find((f) => path.startsWith(`${f.folder}/`));
+			if (home === undefined) {
+				continue;
+			}
+			const name = path.slice(home.folder.length + 1);
+			// Direct children only: a nested path is not an item body.
+			if (name.includes('/') || !isWorkItemFile(name)) {
+				continue;
+			}
+			// Case-INSENSITIVE to match `isWorkItemFile` above: a `Foo.MD` body must
+			// yield the slug `Foo`, or the sidecar-existence guard below would probe
+			// the wrong path and could clear a gate whose sidecar holds an answer.
+			const slug = name.replace(/\.md$/i, '');
+			if (slug === '') {
+				continue;
+			}
+			const item = `${home.type}:${slug}`;
+			// A sidecar STILL EXISTS for this item (canonical or legacy alias) ⇒ this
+			// is the sidecar-anchored sweep's business, not ours. Skipping keeps the
+			// two enumerations DISJOINT, so an item is never planned twice in one
+			// commit and the answered-sidecar carve-out cannot be bypassed through
+			// this path (an item held for an unapplied human answer keeps its flag).
+			if (
+				sidecarPathCandidates(item).some((c) => pathInCommit(base, c, cwd, env))
+			) {
+				continue;
+			}
+			// CONFIRM against the parsed frontmatter: the grep only shortlisted.
+			const body = catBlob(`${base}:${path}`, cwd, env);
+			if (parseFrontmatter(body).needsAnswers !== true) {
+				continue;
+			}
+			out.staleFlags.push({item, itemPath: path});
+		} catch (err) {
+			out.errors.push({
+				item: path,
+				message: err instanceof Error ? err.message : String(err),
+			});
+		}
+	}
+}
+
+/** What a {@link reconcileTerminalQuestionResidue} pass did. */
+export interface TerminalQuestionDrainResult {
+	/** Items whose sidecar was deleted. */
+	drained: string[];
+	/** Items whose `needsAnswers` flag was additionally cleared. */
+	unflagged: string[];
+	/** Terminal items left alone because a human's answer is unapplied. */
+	answeredHeld: string[];
+	errors: {item: string; message: string}[];
+}
+
+/**
+ * Drain the stranded question state {@link classifyTerminalQuestionResidue}
+ * finds, in ONE tree-less commit CAS-published to the arbiter's `main` through
+ * the SAME {@link runTreelessLedgerMove} core the surface path uses (same
+ * contention-retry, same lease, same write seam; there is no second mechanism).
+ *
+ * WHAT IT CLEARS, and the deliberate asymmetry between the two terminals. Note
+ * the terminal SET first: `specs/tasked/` is deliberately NOT in this map at all
+ * (see {@link terminalMainPathsByKind}), so nothing below applies to a tasked
+ * spec, whose question state stays untouched in both halves.
+ *   - the SIDECAR is deleted for EITHER terminal in the map. A question asking
+ *     whether to cancel an item that has already come to rest is stale in both
+ *     cases, and it sits in a folder a human scans carrying a destructive
+ *     default.
+ *   - the `needsAnswers` FLAG is cleared ONLY for a `completed` terminal
+ *     (`tasks/done/`). On a `wont-proceed` terminal
+ *     (`tasks/cancelled/`, `specs/dropped/`) the flag is KEPT, because an item
+ *     can be cancelled precisely BECAUSE its questions were never answered: there
+ *     the flag is accurate history, not residue, and the body may carry a real
+ *     `## Open questions` section saying so. Keeping it is harmless, a terminal
+ *     item is in no pool, so the flag gates nothing.
+ *   - a SUCCESS-terminal item whose gate is armed with NO sidecar left beside it
+ *     has that FLAG cleared and nothing deleted (there is nothing to delete).
+ *     Restricted to the `completed` terminal by the same asymmetry above.
+ *
+ * A sidecar with ANY answered entry is never touched (see the classifier), and
+ * an item still holding such a sidecar is excluded from the flag-only half too,
+ * so the carve-out cannot be bypassed by clearing its gate.
+ *
+ * Best-effort: it never throws, and any fault leaves the state exactly as it was.
+ */
+export async function reconcileTerminalQuestionResidue(params: {
+	cwd: string;
+	arbiter: string;
+	mainRef: string;
+	env?: NodeJS.ProcessEnv;
+	/** Skip the `mainRef` refresh because the CALLER just did it (the combined
+	 * pass refreshes once and runs both sub-passes against that ONE snapshot). */
+	mainAlreadyFresh?: boolean;
+	note?: (message: string) => void;
+}): Promise<TerminalQuestionDrainResult> {
+	const {cwd, arbiter, mainRef, env} = params;
+	const note = params.note ?? (() => {});
+	const result: TerminalQuestionDrainResult = {
+		drained: [],
+		unflagged: [],
+		answeredHeld: [],
+		errors: [],
+	};
+	let report: TerminalQuestionReport;
+	try {
+		report = await classifyTerminalQuestionResidue({
+			cwd,
+			arbiter,
+			mainRef,
+			env,
+			mainAlreadyFresh: params.mainAlreadyFresh,
+		});
+	} catch (err) {
+		result.errors.push({
+			item: '(classify)',
+			message: err instanceof Error ? err.message : String(err),
+		});
+		return result;
+	}
+	result.answeredHeld = report.answeredHeld.map((r) => r.item);
+	result.errors.push(...report.errors);
+	if (report.drainable.length === 0 && report.staleFlags.length === 0) {
+		return result;
+	}
+	// What the LANDED commit ACTUALLY did, filled in by the plan against the base
+	// it committed on. The pre-plan `report` above is only a fast "is there
+	// anything to do?" probe; reporting from it would claim a gate was disarmed
+	// when a contention retry re-derived the residue and skipped the item.
+	let applied: TerminalQuestionResidue[] = [];
+	// Filled by the PLAN with what it actually STAGED (not what it intended), so a
+	// body the marker writer cannot annotate is never reported as unflagged.
+	const clearedSidecarFlags: TerminalQuestionResidue[] = [];
+	const clearedStaleFlags: TerminalFlagResidue[] = [];
+	// NEVER THROW. `runTreelessLedgerMove` and the git plumbing inside the plan
+	// both throw on any non-zero git, and this pass runs from the CLAIM path as
+	// OPPORTUNISTIC HYGIENE on unrelated items. A fault here (a stale scratch ref,
+	// a protected `main`, a permission refusal) must degrade to "left it alone",
+	// never fail the caller's actual work.
+	let landed = false;
+	try {
+		landed = await runTreelessLedgerMove({
+			cwd,
+			// The ref name only has to be unique for the scratch ref; this pass is
+			// batch (many items, one commit), so it is not keyed to a single slug.
+			slug: 'terminal-question-drain',
+			arbiter,
+			kind: 'needs-attention',
+			onContended: 'drain stranded questions',
+			explicitMainRefspec: true,
+			env,
+			note,
+			// RE-PLANNED per attempt against the freshly-fetched base: the residue is
+			// RE-DERIVED from that base, never reused from the probe above, so an item
+			// re-opened out of its terminal folder, or a sidecar a human answered, in the
+			// contention window is correctly left alone.
+			plan: (base) => {
+				const fresh = deriveTerminalQuestionResidue(base, cwd, env);
+				applied = fresh.drainable;
+				result.answeredHeld = fresh.answeredHeld.map((r) => r.item);
+				return prepareTerminalQuestionDrainCommit({
+					cwd,
+					base,
+					residue: fresh.drainable,
+					staleFlags: fresh.staleFlags,
+					clearedSidecarFlags,
+					clearedStaleFlags,
+					env,
+				});
+			},
+		});
+	} catch (err) {
+		result.errors.push({
+			item: '(publish)',
+			message: err instanceof Error ? err.message : String(err),
+		});
+		return result;
+	}
+	if (!landed) {
+		result.errors.push({
+			item: '(publish)',
+			message:
+				'the stranded-question drain did not land on the arbiter’s main ' +
+				'(contention exhausted, or nothing to do); state left untouched.',
+		});
+		return result;
+	}
+	for (const r of applied) {
+		result.drained.push(r.item);
+	}
+	// `unflagged` reports what the commit ACTUALLY staged, from both halves. The
+	// flag-only half never appears in `drained`: it deletes nothing.
+	for (const r of [...clearedSidecarFlags, ...clearedStaleFlags]) {
+		result.unflagged.push(r.item);
+	}
+	return result;
+}
+
+/**
+ * Stage `itemPath` with `needsAnswers` cleared, into the scratch index the drain
+ * commit is being built in. Shared by BOTH halves of the residue (the
+ * sidecar-paired flag and the stranded flag-only one) so they can never disagree
+ * about what clearing a gate means.
+ *
+ * Defense-in-depth, mirroring the surface path's guard in the opposite
+ * direction: if the marker does not parse back as `false`, the body is left
+ * ALONE rather than written as something we cannot vouch for. Every uncertainty
+ * resolves to LEAVING STATE ALONE.
+ *
+ * RETURNS whether the gate was actually STAGED, so callers report EFFECT rather
+ * than INTENT. That distinction is load-bearing here: a body this cannot
+ * annotate (e.g. duplicate `needsAnswers` keys, where the writer replaces the
+ * FIRST and the parser reads the LAST) would otherwise be reported as unflagged
+ * on every claim for ever while its gate stayed armed, the precise
+ * "reports success while the defect remains" failure this whole change exists to
+ * correct.
+ */
+function clearNeedsAnswersInIndex(
+	itemPath: string,
+	base: string,
+	cwd: string,
+	env: NodeJS.ProcessEnv | undefined,
+	withIndex: NodeJS.ProcessEnv,
+): boolean {
+	if (!pathInCommit(base, itemPath, cwd, env)) {
+		return false;
+	}
+	const body = catBlob(`${base}:${itemPath}`, cwd, env);
+	const cleared = setNeedsAnswersMarker(body, false);
+	if (parseFrontmatter(cleared).needsAnswers !== false) {
+		return false;
+	}
+	const blob = hashObject(cleared, cwd, env);
+	gitHard(
+		['update-index', '--add', '--cacheinfo', `100644,${blob},${itemPath}`],
+		cwd,
+		withIndex,
+	);
+	return true;
+}
+
+/**
+ * Build the ONE tree-less commit that removes every drainable sidecar and clears
+ * the `needsAnswers` flag on every `completed`-terminal body, using PLUMBING on a
+ * SCRATCH INDEX (the caller's index/HEAD/working tree are never touched)
+ * exactly as {@link prepareTreelessSurfaceCommit} does in the opposite direction.
+ *
+ * Batched into a single commit on purpose: the residue is a SET, one commit is
+ * one CAS against `main` instead of N, and the whole drain then lands or does not
+ * land atomically.
+ */
+function prepareTerminalQuestionDrainCommit(params: {
+	cwd: string;
+	base: string;
+	residue: TerminalQuestionResidue[];
+	staleFlags: TerminalFlagResidue[];
+	/**
+	 * OUT-PARAM: filled with the items whose gate was ACTUALLY staged as cleared,
+	 * so the caller reports EFFECT rather than intent. Cleared on entry, because a
+	 * contention retry re-plans against a fresh base and the previous attempt's
+	 * result must not leak into the report.
+	 */
+	clearedSidecarFlags: TerminalQuestionResidue[];
+	clearedStaleFlags: TerminalFlagResidue[];
+	env: NodeJS.ProcessEnv | undefined;
+}): TreelessAttemptPlan {
+	const {cwd, base, residue, env, clearedSidecarFlags, clearedStaleFlags} =
+		params;
+	clearedSidecarFlags.length = 0;
+	clearedStaleFlags.length = 0;
+	// RE-DERIVE against THIS base: anything already gone is not our business.
+	const live = residue.filter((r) =>
+		pathInCommit(base, r.sidecarPath, cwd, env),
+	);
+	const liveFlags = params.staleFlags.filter((r) =>
+		pathInCommit(base, r.itemPath, cwd, env),
+	);
+	if (live.length === 0 && liveFlags.length === 0) {
+		return 'already-done';
+	}
+	const scratchIndex = join(
+		tmpdir(),
+		`dorfl-question-drain-${process.pid}-${Date.now()}.index`,
+	);
+	const withIndex: NodeJS.ProcessEnv = {
+		...(env ?? process.env),
+		GIT_INDEX_FILE: scratchIndex,
+	};
+	try {
+		gitHard(['read-tree', base], cwd, withIndex);
+		for (const r of live) {
+			// Remove the stale sidecar.
+			gitHard(
+				['update-index', '--force-remove', r.sidecarPath],
+				cwd,
+				withIndex,
+			);
+			// Clear the flag ONLY on a `completed` terminal (see the doc above).
+			if (r.terminal !== 'completed' || !r.flagged) {
+				continue;
+			}
+			if (clearNeedsAnswersInIndex(r.itemPath, base, cwd, env, withIndex)) {
+				clearedSidecarFlags.push(r);
+			}
+		}
+		// The FLAG-ONLY half: a SUCCESS terminal whose sidecar is already gone. No
+		// `--force-remove` here, because there is nothing to delete; the armed gate IS the
+		// whole residue.
+		// Record what was actually STAGED: a body we could not annotate is dropped
+		// from the report rather than claimed as cleared.
+		for (const r of liveFlags) {
+			if (clearNeedsAnswersInIndex(r.itemPath, base, cwd, env, withIndex)) {
+				clearedStaleFlags.push(r);
+			}
+		}
+		const tree = runHard(['write-tree'], cwd, withIndex).stdout.trim();
+		const touched = live.length + liveFlags.length;
+		const only = live[0]?.item ?? liveFlags[0]?.item;
+		const subject =
+			touched === 1
+				? `drain stranded question state for ${only} (terminal on main)`
+				: `drain stranded question state for ${touched} terminal items`;
+		const commit = runHard(
+			['commit-tree', tree, '-p', base, '-m', subject],
+			cwd,
+			env,
+		).stdout.trim();
+		const ref = 'refs/dorfl/question-drain/batch';
+		gitHard(['update-ref', ref, commit], cwd, env);
+		return {ref, commit};
+	} finally {
+		rmSync(scratchIndex, {force: true});
+	}
+}
+
 export interface SurfaceStuckToNeedsAttentionOptions {
 	/**
 	 * The working clone the move is ORIGINATED from — purely the ORIGIN SOURCE