axstack 0.19.1 → 0.20.1

package/skills/axstack/references/candidate-publication.md

@@ -32,8 +32,8 @@ Any author repair creates a new revision and repeats this boundary.
 
 ## Automation repair exception
 
-For an automation repair under
-[Automation sessions](automations.md), the candidate is a local immutable
+For a manager PR repair under
+[Native PR managers](automations.md), the candidate is a local immutable
 commit SHA in the per-PR child worktree, not a published remote ref. The
 reviewer confirms that exact local SHA with `git rev-parse` in the worktree
 instead of remote equality, and inspects a detached checkout of it with the
@@ -42,4 +42,4 @@ it as the expected-old remote SHA rather than requiring it to equal the
 candidate. Remote equality is re-checked at the publication readback of the
 watch skill's repair-publication reference immediately before the fast-forward
 push. Ordinary workflows keep the remote confirmation above; the
-exception never applies outside an automation session.
+exception never applies outside a manager PR job.

package/skills/axstack/references/lifecycle.md

@@ -17,7 +17,8 @@ binding state and receipts to exact revisions.
   `axstack-reviewer-secondary` sessions with identical brief and isolated first
   pass; authored = one eligible configured reviewer from actual author
   provenance. Owner and author never review.
-- Automation driver/monitor/watchdog: see [Watch health](#watch-health).
+- Automation review/watch managers and optional monitor: see
+  [Watch health](#watch-health).
 - Auditor (`axstack-auditor`): report-only; never edits, merges, activates, or
   audits itself.
 
@@ -109,12 +110,11 @@ differs from merged; human merges.
 
 ## Watch health
 
-The user lifted the native-watch hold by user decision on 2026-09-16. The
-driver automation is a mutating owner for its PRs; `axstack-watchdog` stays
-independent and read-only. The driver every 15 minutes dispatches and exits;
-the watchdog is model-free, has no gate, and records `watchdog.log`; there is
-no watch deadline for automations. Build no custom scheduler and use no legacy
-fallback. Details live in
+Native review and watch lanes use fresh finite sessions in persistent dedicated
+workspaces on staggered 15-minute schedules. Each admits at most five executing PR jobs;
+waiting PRs remain covered without reserving slots. Bounded PR jobs own their
+events, use per-PR worktrees, and settle after descendants settle. Build no
+custom scheduler, state engine, or legacy fallback. Details live in
 [Watch runtime](../../axstack-watch/references/watch-runtime.md).
 
 ## Audit hook (close-out and meaningful checkpoints)

package/skills/axstack/references/orca-runtime.md

@@ -26,7 +26,7 @@ Read `roles.json` from the installed shared root `skills/axstack/`. The installe
 shape is `{ "version": 1, "preset": "<name>", "roles": [...] }`. Bundled
 profiles are setup inputs shaped as
 `{ "version": 1, "roles": [...] }`. A new run records the selected preset and
-all 25 role rows once. An active run keeps the exact snapshot until the user
+all 24 role rows once. An active run keeps the exact snapshot until the user
 explicitly changes it.
 
 Select the requested role by stable ID. A missing or null model holds only that role;

package/skills/axstack/references/review-manager-prompt.md

@@ -0,0 +1,10 @@
+# Review manager prompt
+
+You are a fresh finite review-manager session in its persistent dedicated Orca workspace. Enter
+through the installed `axstack-review` skill; that skill loads its packaged
+`../axstack/references/automations.md` contract by relative reference. Discover
+every eligible peer-review event and invoke the skill through Orca for each
+admitted bounded PR job. Keep complete coverage separate from the five-job
+execution cap, preserve exact ownership and receipts, settle completed trees,
+and stay quiet when nothing changed. Reconcile before admission, save durable
+continuity, then self-close as the final action. Never check out a PR branch here.

package/skills/axstack/references/routing.md

@@ -11,7 +11,7 @@ only with exactly one unambiguous preset; missing or contradictory sources are
 a setup gap: hold. Never infer from live profiles or `list_profiles`, harness,
 tools, credentials, quota, subscription, or default to `mixed`.
 
-At run start, capture one **routing snapshot**: the complete map of all 25 role
+At run start, capture one **routing snapshot**: the complete map of all 24 role
 IDs with provider/model/mode/effort, absent or unconfigured roles recorded
 explicitly, and no invented provider default. An absent or unconfigured role
 holds only that role's work, not the run. A role installed or changed later
@@ -43,8 +43,8 @@ Role IDs:
   `axstack-arena-judge-fable` judge them. `axstack-auditor` audits;
   `axstack-checker` reports discrepancies.
 - `axstack-explainer` authors explanations; `axstack-explainer-review`
-  reviews them. `axstack-monitor` observes only; `axstack-watchdog` sends
-  only gate-authorized health escalations.
+  reviews them. `axstack-monitor` is an optional read-only standalone-watch
+  observer that never sends.
 - `axstack-debug-investigator-1..4` each probe one L1 brief.
 
 Provenance is matched on provider/model ID; effort never maps. Missing table-row

package/skills/axstack/references/watch-manager-prompt.md

@@ -0,0 +1,11 @@
+# Watch manager prompt
+
+You are a fresh finite watch-manager session in its persistent dedicated Orca workspace. Enter
+through the installed `axstack-watch` skill; that skill loads its packaged
+`../axstack/references/automations.md` contract by relative reference. Discover
+every eligible own-PR event and invoke the skill through Orca for each admitted
+bounded PR job. Keep complete coverage separate from the five-job execution
+cap, preserve exact ownership, candidates, holds, and receipts, settle
+completed trees, and stay quiet when nothing changed. Reconcile before
+admission, save durable continuity, then self-close as the final action. Never
+check out a PR branch here.

package/skills/axstack-relay/SKILL.md

@@ -64,9 +64,9 @@ listing all pass.
 ## Preserve identity and authority
 
 Delivery is one-way; no session polls Telegram. Hermes does not route a reply
-back to the sending session; its own agent answers replies. Outside the fixed
-decision-token flow below, a reply is never a receipt, decision, or authority
-for this session, and no persistent owner is needed to send. Every ordinary
+back to the sending session; its own agent answers replies. A reply is never a
+receipt, decision, or authority for this session, and no persistent owner is
+needed to send. Every ordinary
 message must say where the user acts: the current Orca conversation, the Orca
 worktree, or the GitHub PR. Do not invent reply commands.
 
@@ -76,25 +76,6 @@ authority. Delivery is transport evidence only. Revalidate any user decision
 that arrives through an authorized channel against the current task and
 existing action boundaries before acting; silence never grants permission.
 
-## Decision tokens
-
-An automation escalation is the narrow exception defined by
-[Automation sessions](../axstack/references/automations.md). The PR agent opens
-an immutable-bound decision token, sends one message with the exact
-`/axstack-decide approve <token>` and `/axstack-decide reject <token>` replies,
-records the send receipt, and
-exits without waiting. The Hermes script decides by validating the private
-channel and updating only an open token file; Hermes never performs the bound
-GitHub or Git action. The driver consumes the file on a later tick, revalidates
-all bound state, marks an approval spent before acting, performs only that
-action, and records its receipt.
-
-This does not create a reply channel for the sending agent: Hermes pushes the
-decision to the file through the fixed `axstack-decide` script, and the driver
-reads that file. Delivery is one-way; no session polls Telegram. Token creation,
-decision, and consumption keep their separate writers and authority; a message
-receipt alone authorizes nothing.
-
 ## Reconcile, deliver, and record
 
 Before any new send, check the caller's run record for an existing receipt

package/skills/axstack-review/SKILL.md

@@ -9,6 +9,12 @@ Produce one evidence-bound verdict for an exact candidate revision using the
 review count and model routing required by its mode. Report within the
 requested authority; the human merges unless separately authorized otherwise.
 
+When the current session is a fresh review-manager session, load
+[Native PR managers](../axstack/references/automations.md) and follow only its
+discovery, admission, recovery, and settlement branch. Do not review a PR,
+materialize `axstack-owner`, or check out a PR branch in the manager workspace.
+Each admitted bounded PR coordinator re-enters this skill in peer mode.
+
 Before reviewing, load [Standing contracts](../axstack/references/contracts.md),
 then [Lifecycle and receipts](../axstack/references/lifecycle.md) so its required
 audit edge remains active. Load [Shared routing](../axstack/references/routing.md)
@@ -16,8 +22,8 @@ to select the mode and scope identity, and apply the shared
 [PR-shape policy](../axstack/references/pr-shape.md). For an owned candidate,
 load and verify the
 [candidate-publication boundary](../axstack/references/candidate-publication.md).
-When the caller is an Orca driver automation, load
-[Automation sessions](../axstack/references/automations.md): its reviewer briefs
+When the caller is a bounded review-manager PR job, load
+[Native PR managers](../axstack/references/automations.md): its reviewer briefs
 carry the required escalation field and every eligible peer PR takes a binding
 `APPROVE` or `REQUEST_CHANGES` verdict under the automation exception below.
 
@@ -71,13 +77,13 @@ fallback.
 Before dispatch, read [Orca runtime](../axstack/references/orca-runtime.md).
 Standalone peer review or watch adoption then materializes `axstack-owner`,
 reusing a live owner when one exists. Once materialized, that owner is the sole
-coordinator: only the owner launches the writer, reviewers, monitor, and
-watchdog. The current chat does not compete with it. Workers create no children
-or recursive teams.
+coordinator: only the owner launches the writer, reviewers, and optional
+monitor. The current chat does not compete with it. Leaf workers create no
+recursive teams.
 
 Automation exception — Standalone owner: no separate `axstack-owner` is
-materialized when the caller is the Orca driver automation; the automation
-session is the owner for every PR it handles.
+materialized when the caller is a bounded manager PR job; that PR coordinator
+owns the event and settles after its skill-owned reviewers settle.
 
 ## Review the candidate
 
@@ -230,9 +236,9 @@ Escalate to user: yes | no — <criterion> — <reason>
 Every brief ends with the `Escalate to user` field and the reviewer answers it
 in the receipt. A reviewer may cite only a security concern, a permanent
 on-chain state change, or an architectural change in approach. Health is not a
-reviewer criterion. The answer is input to the PR escalation gate, not a veto
-and not a verdict; see
-[Automation sessions](../axstack/references/automations.md) for the gate.
+reviewer criterion. The answer is escalation input, not a veto or verdict; see
+[Native PR managers](../axstack/references/automations.md) for the manager-chat
+hold.
 
 ## Template: review receipt (one block per revision)
 
@@ -268,12 +274,11 @@ relay delivery fails, send the same escalation there. Failed delivery never reso
 concern. Use no private escalation script. Public installations inherit no
 private transport values or configuration.
 
-Under an automation session, credible serious risk found by a reviewer still
-raises the standing internal prompt and dependent-action hold immediately, and
-the gate governs only external notification: the internal prompt lands in the
-run record and the automation session's own Orca conversation. `escalate`
-opens the bound decision token, sends through `axstack-relay`, and exits
-without waiting; `proceed` never overrides a validated blocking finding.
+Under a manager PR job, credible serious risk found by a reviewer raises the
+standing internal prompt and dependent-action hold immediately in the compact
+run record and a durable GitHub or user-owned conversation. The authorized
+`axstack-relay` notification points the user there; delivery or silence never
+authorizes action.
 
 ## Publishing rule
 
@@ -322,13 +327,12 @@ to the intended commit.
 ## Automation exception
 
 For a peer PR selected under
-[Automation sessions](../axstack/references/automations.md), apply the same
-complete-review and exact-commit requirements, then run the Luna gate.
-`escalate` opens a bound decision token, sends its message, and exits without
-submitting. `proceed` permits `APPROVE` only with no validated blocker and
-permits `REQUEST_CHANGES` only with at least one evidenced validated blocker.
-`INCOMPLETE`, unavailable inputs, unresolved disagreement, or unknown GitHub
-state submits nothing.
+[Native PR managers](../axstack/references/automations.md), apply the same
+complete-review and exact-commit requirements. A serious-risk escalation holds
+submission at its durable decision location. Otherwise `APPROVE` requires
+no validated blocker and `REQUEST_CHANGES` requires at least one evidenced
+validated blocker. `INCOMPLETE`, unavailable inputs, unresolved disagreement,
+or unknown GitHub state submits nothing.
 
 Immediately before `gh pr review`, re-read self's reviews at the head. If one
 already exists, skip submission and record its id. Otherwise re-check head,

package/skills/axstack-watch/SKILL.md

@@ -14,6 +14,13 @@ Its required edge loads [Shared lifecycle](../axstack/references/lifecycle.md),
 including the end-of-run audit hook. Reach other references only at the steps
 that name them.
 
+When the current session is a fresh watch-manager session, load
+[Native PR managers](../axstack/references/automations.md) and follow only its
+discovery, admission, recovery, and settlement branch. Do not adopt or repair a
+PR, materialize `axstack-owner`, or check out a PR branch in the manager
+workspace. Each admitted bounded PR coordinator re-enters this skill in its
+recorded mode.
+
 Preserve any explicitly named PR, repository, or peer scope. For broad
 discovery of the user's own PRs (such as “my” or “our” PRs), run
 `gh api user --jq .login` on the execution host, then select open PRs authored
@@ -21,9 +28,9 @@ by that login in the named or current repository. Never hardcode or guess the
 username; a missing or failed authenticated-login lookup is a concrete blocker.
 The authenticated human login selects PRs. Runtime session IDs coordinate work
 only and establish neither human identity nor write, reply, or merge authority.
-When the session is the Orca driver automation, also load
-[Automation sessions](../axstack/references/automations.md): it is the owner
-for every PR it handles, and its allowlist bounds every mutation.
+When the session is a bounded watch-manager PR job, also load
+[Native PR managers](../axstack/references/automations.md): the bounded PR
+coordinator owns that event and its allowlist bounds every mutation.
 
 ## 1. Adopt and reconcile
 
@@ -64,21 +71,20 @@ Read-only checks and updates to the already-owned local record need no runtime
 load. When the watch needs a new owner or automated observation, first read
 [Watch runtime](references/watch-runtime.md) and then
 [Orca runtime](../axstack/references/orca-runtime.md). Reconcile before creating
-anything. The user lifted the native-watch hold by user decision: the driver
-every 15 minutes dispatches and exits as a mutating owner; the watchdog is
-model-free and read-only, has no gate, and records `watchdog.log`; there is no
-watch deadline for automations. The driver is the automation session itself,
-with no `axstack-monitor` or `axstack-owner` role row; `axstack-monitor` stays
-an optional read-only observer that never sends. One read-only PR observation
-needs neither. The publishing driver is the live owner for a status check;
+anything. Each native watch-manager pass starts a fresh finite session in one
+persistent dedicated workspace on its staggered 15-minute schedule, covers
+every eligible own PR, and starts only bounded actionable-event jobs. Waiting PRs reserve no execution slots and
+there is no watch deadline for manager automation. `axstack-monitor` stays an
+optional read-only observer that never sends. One read-only PR observation
+needs neither. The bounded PR coordinator is the live owner for its event;
 materialize no `axstack-owner` and start no automation for a read-only check.
 
 For standalone adoption, materialize `axstack-owner` only when no live owner
 exists. Once it exists, the current chat is not a competing coordinator. Only
-the owner launches the writer, reviewers, monitor, and watchdog. Workers create
-no children or recursive teams, and the adoption watcher is never the writer.
+the owner launches the writer, reviewers, and optional monitor. Leaf workers
+create no recursive teams, and the adoption watcher is never the writer.
 
-A live watch has verified role and timer receipts, handshakes, watched scope,
+A standalone live watch has verified role and timer receipts, handshakes, watched scope,
 wake ownership, and a common expiry. A missing runtime capability is a setup gap,
 not a reason to invent a call or create a duplicate registration. Native
 wake-ups drive observation; never poll or keep a model active between events.
@@ -94,9 +100,9 @@ A healthy unchanged observation produces no user-facing message.
 
 Observation-only and peer wakes produce a read-only report and stop. For an
 authorized maintenance wake that may require a repair or public reply, read and
-follow [Repair and publication](references/repair-publication.md). An
-automation session repairs in a per-PR child worktree created through
-`orca-cli`; its driver worktree never checks out a PR branch.
+follow [Repair and publication](references/repair-publication.md). A manager PR
+job repairs in its per-PR child worktree created through `orca-cli`;
+the manager workspace never checks out a PR branch.
 
 ### Feedback routing
 
@@ -107,9 +113,9 @@ snapshot. Missing, stale, or materially changed identity holds repair routing
 while monitoring continues. Accepted fixes return to the same original author
 session only when the run itself launched that session and evidence allows,
 then receive refreshed review under the authored mode rule before publication.
-For an adopted own PR under the automation, the original authoring session is
-not a run-launched session: the repair author is the automation session
-(Claude/Opus) or a dispatched `axstack-author` (Sol), and the authored-review
+For an adopted own PR under the manager, the original authoring session is
+not a run-launched session: the repair author is the manager PR coordinator or
+a dispatched `axstack-author`, and the authored-review
 pairing follows the recorded actual provenance of that repair, not the PR's
 historical author. Unknown, mixed, or unsupported author provenance
 that cannot establish the eligible configured reviewer is an exact gap to
@@ -120,10 +126,9 @@ the current revision, and a recorded hold or next owner where work remains.
 
 When a new actionable event is eligible under a recorded `Notification policy`,
 the owner may use the optional [axstack-relay](../axstack-relay/SKILL.md).
-The monitor never sends. The automation watchdog never mutates GitHub, has no
-gate, and records occurrences and send receipts in `watchdog.log`; absent
-policy or failed relay uses the current Orca conversation and leaves every
-existing hold open.
+The monitor never sends. The manager deduplicates authorized notifications;
+absent policy or failed relay uses the recorded durable GitHub or user-owned
+conversation and leaves every existing hold open.
 
 ## 5. State readiness precisely
 
@@ -134,9 +139,14 @@ observed state distinct from merged, and the human merges by default.
 
 ## 6. End and preserve continuity
 
+A bounded manager PR job ends as soon as its current event and every owned
+descendant settle. It returns exact receipts and remaining state to the logical
+manager lane's durable continuity, releases proven resources, and never waits
+for merge or stops the manager's recurring schedule.
+
 End a standalone watch early when all required PRs merge, at cancellation, or
-at its shared default 24 h deadline. There is no watch deadline for
-automations. In every case, stop and verify all owned registrations.
+at its shared default 24 h deadline. There is no watch deadline for manager
+automation. In every case, stop and verify all owned registrations.
 
 At every end condition, leave the compact state below in the private run record
 and report it in the current chat, even when work remains. Expiry grants neither

package/skills/axstack-watch/references/repair-publication.md

@@ -8,9 +8,9 @@ and peer modes stop with a report before this branch.
 Re-read the accepted maintenance snapshot, writable ownership, publication
 authority, remote head and base, and current feedback. Route an accepted fix to
 the original author session only when the run itself launched that session and
-evidence permits. For an adopted own PR under the automation, the repair
-author is the automation session (Claude/Opus) or a dispatched
-`axstack-author` (Sol); record that repair's actual provenance before
+evidence permits. For an adopted own PR under the manager, the repair author is
+the manager PR coordinator or a dispatched `axstack-author`; record that
+repair's actual provenance before
 selecting the reviewer, because the authored-review pairing follows the actual
 provenance of the repair, never the PR's historical author. A missing,
 stale, or materially changed boundary holds the repair while read-only
@@ -34,13 +34,11 @@ required by the authored review rule in `axstack-review`.
 Publication stays held until the current authored review receipt covers the
 exact new revision and base, all six angles, applicable acceptance, the reply
 body identities, and every affected boundary, with no unresolved material
-finding or urgent hold. Under an automation the escalation gate of
-[Automation sessions](../../axstack/references/automations.md) also runs on
-the local SHA. `escalate` creates a token, pins the candidate at
-`refs/axstack/decisions/<token>`, opens the decision token, sends its message,
-and exits instead of pushing. `proceed` permits publication only with no
-unresolved validated blocking finding, and a push before the gate settles is
-forbidden.
+finding or urgent hold. Under a manager PR job, a serious-risk escalation
+preserves the local candidate and exact context, records the hold in the
+compact record and a durable GitHub or user-owned conversation, and sends only
+an authorized deduplicated notification. Publication remains held until the
+user decides at that durable location and all exact inputs are revalidated.
 
 ## 3. Revalidate immediately before publication
 
@@ -57,9 +55,9 @@ Use `gh stack` for the adopted PR only, preserving unrelated stack entries.
 Bind the operation to the exact reviewed revision, then verify the submission
 receipt and remote state.
 
-An automation session pushes fast-forward only: run the section 3 publication
+Manager automation pushes fast-forward only: run the section 3 publication
 readback immediately before the push, then `git push` to the PR branch with no
-lease or force, and no `gh stack` sync or restack from an automation. A
+lease or force, and no `gh stack` sync or restack from a manager. A
 non-fast-forward remote is a recorded hold, never a rewrite.
 
 If the send outcome is unknown, inspect remote IDs, bodies, and actor before any

package/skills/axstack-watch/references/watch-runtime.md

@@ -2,53 +2,39 @@
 
 Read this before starting, resuming, or stopping automated PR observation.
 
-## Accepted policy
-
-The PR owner remains accountable throughout a standalone watch's default
-24-hour window. `axstack-monitor` and `axstack-watchdog` are independent,
-read-only roles, not authors, reviewers, repliers, or owners.
-
-- **Monitor:** `axstack-monitor` is an optional read-only observer that reads
-  GitHub, all PR feedback, and latest checks on its registered cadence, persists event
-  IDs, wakes the owner only for a new actionable event, and never sends.
-- **Watchdog:** for the PR automations, it is a model-free hourly shell precheck
-  that never mutates GitHub, has no gate, and records each tick in
-  `watchdog.log`.
-
-Healthy observations are snapshot-only and update quietly; they wake neither owner nor
-driver. Both roles deduplicate event IDs. Uncertain delivery is reconciled
-before retry. Restart reuses prior watch identity rather than registering a
-duplicate. The shared deadline ends earlier on completion or cancellation and
-is never silently renewed.
-
-## Native Orca automations
-
-Load the version-matched Orca automation guidance through the shared
-[runtime boundary](../../axstack/references/orca-runtime.md). The verified
-native automation schema supports provider selection, but model, effort, and
-permission pinning are unsupported, and its schedule parser cannot preserve
-the accepted bounded expiry by itself. Requested role values or a post-launch
-self-report are not effective launch evidence.
-
-The user lifted the native-watch hold by user decision on 2026-09-16. The
-driver every 15 minutes dispatches and exits as the mutating owner, and the
-watchdog is model-free and read-only, has no gate, and records `watchdog.log`;
-there is no watch deadline for automations. The driver is the automation
-session itself, with no `axstack-monitor` or `axstack-owner` role row. Still
-introduce no custom scheduler or polling loop and use no legacy runtime
-fallback. The session-level contract lives in
-[Automation sessions](../../axstack/references/automations.md).
-
-## Preserve the contract under automation
-
-An automation session preserves its 15-minute driver and hourly model-free
-watchdog cadences, quiet healthy behavior, deduplication, handshake, watched
-scope, and wake owner. Test restart, duplicate ticks, cancellation,
-session-reuse fallback, and final cleanup before enabling.
-A firing timestamp proves neither delivery nor work advancement. An unrequested
-fallback session reconciles ownership and never becomes owner silently.
-
-At every end condition, stop all task-owned registrations, verify runtime
-cleanup receipts, and capture remaining work as resumable state. Removing watch
-coverage, weakening role discipline, or changing the deadline requires a
-material specification revision; it is not an implementation workaround.
+## Standalone watch
+
+A standalone PR owner remains accountable through the default 24-hour window.
+`axstack-monitor` is an optional independent read-only observer: it reads the
+current GitHub state, persists event IDs, wakes the owner only for a new
+actionable event, and never sends or mutates. Healthy observations update
+quietly. Reuse prior watch identity rather than registering a duplicate, and
+stop task-owned registrations at completion, cancellation, or expiry.
+
+## Native watch manager
+
+The separate native watch manager lane follows the
+[native PR-manager contract](../../axstack/references/automations.md). It reuses
+one persistent dedicated workspace while each scheduled pass uses a fresh
+finite session at minutes `7,22,37,52`. It scans every eligible own PR and
+admits at most five bounded actionable-event jobs. Waiting PRs stay covered
+without reserving slots. Each job uses its own repository-parented worktree and
+settles with all descendants; no model, per-PR timer, or polling loop waits for
+CI, review, a user decision, or merge.
+
+Load the version-matched Orca automation and orchestration guidance through the
+shared [runtime boundary](../../axstack/references/orca-runtime.md). Requested
+settings or source configuration are not runtime evidence. Preserve quiet
+unchanged behavior, exact-head event dedupe, ownership, complete discovery,
+and current authority.
+
+Before activation, canary fresh-session launch, an overlapping scheduled pass,
+session loss and recovery, nested dispatch depth for coordinator-launched leaves, and
+total process and memory behavior. Until observed, those remain unverified; do
+not replace them with a precheck, watchdog, custom scheduler, state engine, or
+legacy fallback.
+
+On recovery reconcile actual native workers, GitHub, and the compact run record
+before admitting work. Unknown state blocks only the affected PR. At every job
+end, settle owned workers, preserve unsafe or user-owned state, release proven
+resources, and leave exact resumable receipts.