Fix post-apply stand-down and verify installed Herdr integration

README.md

@@ -207,6 +207,11 @@ attached-agent tuple is also required when an agent was observed.
 
 ## Evidence and release boundary
 
+The [2026-09-14 installed-action developer smoke](docs/evidence/2026-09-14-developer-installed-action-smoke.md)
+records a fresh isolated Herdr 0.7.5 run through Stage 3 apply and stand-down,
+including the completed-run stand-down fix. It uses deterministic local worker
+fixtures and is separate from formal release evidence and real model integration.
+
 Retained Stage 1 B0/B4 artifacts remain historical compatibility evidence. The
 [B4 live smoke report](docs/evidence/2026-07-28-stage1-b4-live-smoke.md) is
 validated against its recorded candidate Git tree, not current Stage 2 bytes.

docs/evidence/2026-09-14-developer-installed-action-smoke.md

@@ -0,0 +1,105 @@
+# Installed-action developer smoke — 2026-09-14 UTC
+
+The installed Conductor lifecycle passed on an isolated, running Herdr 0.7.5
+server after fixing argument-free stand-down following Stage 3 apply.
+
+This is bounded developer smoke evidence, not the formal Stage 2 result trio,
+an independent-human source review, release attestation, or production approval.
+Historical evidence and its review requirements remain unchanged.
+
+## Tested source and environment
+
+| Field | Observed value |
+| --- | --- |
+| Corrected candidate | `0f0e907742292b733d3bb8cf569096fbe265beba` |
+| Plugin version | `0.4.0` (unreleased fix; no version bump) |
+| Herdr client and server | `0.7.5` |
+| Server protocol / API schema | `17` / `1` |
+| Platform | macOS, Apple Silicon |
+| Installed-action and report-publisher Node | `24.18.0` |
+| Attended approval-recorder Node | `26.4.0` |
+| Official runtime asset | [herdr-macos-aarch64, v0.7.5](https://github.com/herdrdev/herdr/releases/tag/v0.7.5) |
+| Runtime asset SHA-256 | `37350546b0012555943b92eaf962665de4e264395baeb44227b8015e8ff5b0d6` |
+
+The candidate was cloned at the exact commit and remained clean. A private
+temporary root contained a dedicated API socket, XDG config/state directories,
+plugin registry, synthetic Git repositories, and explicit absolute Conductor
+state. No default session or production repository was used. The synthetic
+repositories had no remotes.
+
+A local shell fixture occupied the Pi worker slots and reported synthetic
+session identities using Herdr's `herdr:pi` integration protocol. It invoked no
+model and read no agent credentials. The operator made the deterministic change
+and inspected the validator snapshot, then published reports through the actual
+candidate report publisher. This tests installed plugin actions and real Herdr
+transport; it does not validate a real Pi/model integration or autonomous work.
+
+## Failure found and correction
+
+Candidate `bf67d318067c60318dcd7089897f513bfefe3af7` successfully previewed,
+approved, applied, and replayed the fixture result, but its installed stand-down
+action failed with `bookkeeping_unknown: stand-down reason does not match
+lifecycle state`. No worker panes were closed by that failed action.
+
+The action supplies no reason. The runtime defaulted to `operator_abandoned`,
+while an applied run correctly permits `normal_completion` only. The corrected
+candidate selects `normal_completion` for an applied run when the reason is
+omitted. Explicit invalid reasons, other lifecycle defaults, recorded restart
+authority, and archive replay retain their existing checks. A regression failed
+before this change and passed afterward.
+
+Initial fixture setup also exposed the expected refusal when no recognized
+worker session identity was present. Those failed runs were retained; a fresh
+repository/run was used after correcting the fixture's session reporting. No
+journal or run authority was rewritten to advance a failed run.
+
+## Observed corrected workflow
+
+1. Linked the exact candidate in the isolated registry and confirmed all seven
+   installed actions. Committed configuration v3 with one builder, one validator,
+   and an existing unchecked-out `refs/heads/live-apply-target` at the fork SHA.
+2. Invoked installed assemble, board, and status. The builder's real worktree,
+   pane, task, outbox, and session identity were observed.
+3. Committed `src/smoke.txt` containing `verified local fixture` plus a newline,
+   published the builder report, and invoked harvest. Integration performed one
+   ref compare-and-swap and dispatched the validator at the exact final SHA.
+4. Checked the validator's file content and detached HEAD against its task's
+   integration SHA, published its report, and harvested it successfully.
+5. Invoked preview and inspected the exact target, one-file diff, and accepted
+   validator assertion. The target remained at its original SHA. Apply without
+   a receipt and a malformed receipt were refused.
+6. The attending operator recorded a canonical approval for that exact preview,
+   then invoked installed apply. One approval consumption preceded one apply
+   publication in the durable journal; the target moved once to the final SHA.
+7. Repeated apply. It returned the recorded outcome with `replayed: true`.
+   There remained one consumption and one publication, and the target reflog
+   contained only its branch-creation entry and the single apply update.
+8. Invoked installed stand-down with no reason. It archived the run and closed
+   exactly the builder and validator panes, leaving the invoking root pane.
+   Repeated stand-down returned the archived result with `replayed: true`.
+
+| Final observation | Value |
+| --- | --- |
+| Apply target before | `6e67b667f187651a21826fdf5e255128dc9f4e70` |
+| Apply target after | `30e7ac2ebda9fcdc34883aeaa704c05a916a315e` |
+| Fixture file SHA-256 | `77a2cdc00810ab5cbd0be50c7895f7af8bfd7b77a2d6d24a3450e87883311685` |
+| Approval consumption / publication sequence | `16` / `17` |
+| Apply CAS count | `1` |
+| Retained Git worktrees after stand-down | `3` |
+| Worker panes closed by stand-down | `2` |
+
+Tasks, reports, journals, branches, and worktrees remained retained. The isolated
+plugin was unlinked and its dedicated server stopped; process exit and socket
+removal were checked. Temporary fixture resources were retained privately for
+inspection. Raw paths, session identifiers, pane output, and private state are
+not included in this public record.
+
+## Coverage limits
+
+No newer-Herdr compatibility, real model-agent integration, broad production
+readiness, target-drift/rejection matrix, or live crash-recovery claim follows
+from this smoke. Crash boundaries and the broader refusal matrix remain covered
+by automated tests. These observations and synthetic worker reports are not
+authenticated. The formal evidence runner and finalizer were not invoked, and
+no independent-human GO record or historical result trio was fabricated or
+replaced.

docs/readiness.md

@@ -42,8 +42,11 @@ the independent exact-source review described in [CONTRIBUTING](../CONTRIBUTING.
 
 Priorities for subsequent work:
 
-- Obtain candidate-specific live evidence for Stage 3 before making broader
-  deployment claims; preserve the separate independent review requirement.
+- The [installed-action developer smoke](evidence/2026-09-14-developer-installed-action-smoke.md)
+  covers Stage 3 preview, attended apply, replay, and completed stand-down on
+  exact Herdr 0.7.5 with deterministic local workers. Validate real agent
+  integrations before broader deployment claims; preserve the separate formal
+  evidence and independent review requirements.
 - Test additional Herdr versions explicitly before widening the exact runtime
   contract or the website's compatibility claims.
 - Improve operator guidance for retained resources and uncertain pane/agent

scripts/stage1-runtime.mjs

@@ -2044,7 +2044,9 @@ export async function standDown({
 					"stand-down restart reason differs from observed authority",
 				);
 		} else {
-			const selectedReason = reason ?? "operator_abandoned";
+			const selectedReason =
+				reason ??
+				(lifecycle.state === "applied" ? "normal_completion" : "operator_abandoned");
 			standDownReasonForState(lifecycle.state, selectedReason);
 			standDownIdentity = {
 				document_type: "herdr-conductor-stage2-stand-down",

tests/stage3-actions.test.mjs

@@ -22,6 +22,7 @@ import {
 	git,
 	publishWorkerReport,
 	readStatus,
+	readJournals,
 	recordStage3Receipt as record,
 	stage3ApplyOverrides,
 	stage3HarvestedApplyFixture,
@@ -46,6 +47,46 @@ function applyCasCount(fixture) {
 	).length;
 }
 
+test("argument-free stand-down completes an applied run without permitting abandonment", async () => {
+	const { fixture, harvested } = await harvestedApplyFixture(31);
+	const previewed = await preview(invocation(fixture));
+	await record(fixture, receiptFor(previewed, "approve"));
+	await applyStage3(invocation(fixture));
+	const beforeRef = git(fixture.repository, "rev-parse", APPLY_TARGET);
+	const effectsBefore = fixture.fake.effects;
+	await expectCodeAsync("bookkeeping_unknown", () =>
+		standDown(invocation(fixture, { reason: "operator_abandoned" })),
+	);
+	assert.equal(fixture.fake.effects, effectsBefore);
+	assert.equal(
+		readJournals(fixture.stateRoot).filter(
+			(entry) => entry.operation_type === "run.stand-down.begin",
+		).length,
+		0,
+	);
+	const stood = await standDown(invocation(fixture));
+	assert.equal(stood.archived, true);
+	assert.deepEqual(
+		stood.closed,
+		fixture.result.workers.map(({ pane_id }) => pane_id),
+	);
+	const begin = readJournals(fixture.stateRoot).find(
+		(entry) =>
+			entry.operation_type === "run.stand-down.begin" &&
+			entry.phase === "observed",
+	);
+	assert.equal(begin.observed_identity.source_state, "applied");
+	assert.equal(begin.observed_identity.reason, "normal_completion");
+	assert.equal(begin.observed_identity.outcome, "completed");
+	assert.equal(beforeRef, harvested.integration.final_sha);
+	assert.equal(git(fixture.repository, "rev-parse", APPLY_TARGET), beforeRef);
+	const effectsAfter = fixture.fake.effects;
+	const replayed = await standDown(invocation(fixture));
+	assert.equal(replayed.archived, true);
+	assert.equal(replayed.replayed, true);
+	assert.equal(fixture.fake.effects, effectsAfter);
+});
+
 test("preview, approval, consumption, and apply move the target exactly once", async () => {
 	const { fixture, harvested } = await harvestedApplyFixture(1);
 	const effectsBeforePreview = fixture.fake.effects;