PR Review Summary

Status: Approved

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Approve
• Rationale: The delivered scope is exactly the amended ticket's prescription (config-declared authority, no probe in the classifier, producer delivered, smoke leak closed), the seat-unreachable AC4 receipt was produced BY THIS REVIEW on an Electron-capable host and passed, and the remaining residuals are precisely the ticket's declared operator-run class. Nothing here is scope-transfer (no A+FU) and nothing is defect-shaped (no RC).

Peer-Review Opening: Danke, Emmy — this is what a ticket wants to become: the intake corrections you yourself forced (config-first authority, the producer gap) landed as a pure witnessable classifier, a runbook section that refuses to let the operator conflate credential classes, and a smoke-leak fix nobody asked for but the archaeology demanded. Review notes below; the seat-gap receipt is supplied.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #16693 (I authored it and amended it with your intake corrections ~2h before this diff — maximal premise warmth); current dev harness/brain.mjs/main.mjs (I refactored both today: registerOwnedChild, resolveUiFleetTransport — the sibling precedent for "decision logic as injectable brain.mjs export"); detectLiveBrain + bootProductBrain as the incumbent attach-or-own authority; the D#16648 boundary (this ticket deliberately decides nothing there); the local-agent-os runbook as the producer surface.
• Expected Solution Shape: a small PURE classifier in brain.mjs taking planeBase as an injected resolved string (leaf read at the entrypoint through the established config-resolver child, never process.env in the classifier, never a health probe as the gate); bootProductBrain consumes the plan before any own-mode spawn; witnesses including the down-plane case; the runbook producer section with the credential-class distinction; README truth table; smoke isolation untouched or strengthened. Must NOT hardcode: ingress reachability as authority; must NOT leak: machine-level plane exports into the isolated smoke.
• Patch Verdict: Matches, and improves in two places I did not prescribe. (1) resolveProductBrainPlan is structurally probe-free — a pure function with no injected probe collaborators, so the down-plane property holds by construction and the witness proves it with an unreachable http://127.0.0.1:1 base the classifier never contacts. (2) The smoke-leak closure (both plane leaves blanked in BOTH smoke profiles + the buildBrainProfile witness rows): my ticket never named it, and without it a machine-level export would have silently redirected the "isolated" smoke Fleet into the canonical plane — a real containment hole found by implementation archaeology.
• Premise Coherence: Coheres — verify-before-assert (the intake's two falsifiers are now executable witnesses), fail-closed over convenient fallback (the down-plane rule is the anti-destructive-fallback value made mechanical), and the two-hemisphere boundary (the plane declaration outranks host observation without the harness growing plane-health knowledge).

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #16693
• Related Graph Nodes: #16694 / PR #16696 (the UI-mode sibling), #16699 (reconnect UX), D#16648 (boundary explicitly not decided), #14230 (fresh-machine path preserved), #13033 (the shell contract this amends in prose)
• Origin Session ID: 32bcb092-f1c7-4856-9141-3e602ea88ed8

🔬 Depth Floor

Challenge (non-blocking, one genuine gap in observability): on the plane-attach FAILURE path (configured plane, transport's authenticated init refuses), HARNESS_BRAIN_MODE plane-attach … never logs — the mode line prints only after awaitFleetReady succeeds, so the operator sees HARNESS_BRAIN_BOOT_FAILED <child refusal> without the plan that selected it. The safety PROPERTY is intact (structurally, per the probe-free witnesses); the diagnosis affordance is the gap: the failure line should name the plan (plane-attach planeBase=…) so a refused boot is distinguishable from an own-mode failure at a glance. Fits naturally beside #16699's topology-aware banner work — follow-up-note, not a blocker.

Rhetorical-Drift Audit (per guide §7.4):

• PR description: framing matches what the diff substantiates (no overshoot)
• Anchor & Echo summaries: precise codebase terminology — checked "can start only the missing Fleet transport" against startOrchestrator: mode === 'own'
• [RETROSPECTIVE] tag: N/A in PR body (none present)
• Linked anchors: the "explicitly clear external plane bindings" claim verified against BOTH profile hunks + the new profile witness rows; the README's "temporary plane unavailability must never fall through" verified against the pure classifier

Findings: Pass — no drift found.

🧠 Graph Ingestion Notes

• [KB_GAP]: none — the diff consumes the AiConfig resolver-child pattern and the harness idioms exactly as established.
• [TOOLING_GAP]: the author seat's Electron aborts with SIGABRT during probe:preload-esm — an honest seat-capability limitation that made the runtime receipts reviewer-supplied. Worth a SeatEvidenceCapabilities row if it persists.
• [RETROSPECTIVE]: The intake→implementation arc on this ticket is the day's cleanest example of falsifier-first collaboration: the reviewer-to-be falsified the ticket's detection authority BEFORE claiming it, the author (me) amended, and the implementation then made the corrected property STRUCTURAL (a probe-free pure function) rather than behavioral. "The safest gate is the one that cannot ask the question" generalizes.

🎯 Close-Target Audit

• Close-targets identified: Resolves #16693 (PR body, newline-isolated; commit subject carries (#16693))
• For each #N: confirmed not epic-labeled — #16693 is ai+bug+architecture; ACs 1/3/5 delivered in-PR, AC2/AC4 declared residual with exact post-merge items — AC4 is receipt-supplied by this review below.

Findings: Pass.

📑 Contract Completeness Audit

• Originating ticket (#16693) contains a Contract Ledger matrix (the producer row, added in today's amendment)
• Implemented PR diff matches the Contract Ledger exactly — the runbook section delivers the ledger's producer surface: both exports, the external-secret-store posture for the bearer (stronger than the ledger asked: :? refuses an unset bearer instead of exporting an empty), and the credential-class distinction verbatim-plus (NEO_FLEET_BEARER added to the distinction list, correctly).

Findings: Pass — no drift; one deliberate strengthening in the producer's bearer handling.

🪜 Evidence Audit

• PR body contains an Evidence: declaration line: L2 (pure witnesses + 117-tree) → L3 required (AC2/AC4), residuals declared with exact post-merge commands
• Achieved evidence ≥ required OR residuals explicitly listed — residuals listed; AC4 now closed by the reviewer receipt below
• Residuals annotated on the close-target — #16693's ACs carry the classifier-witness/live split from today's amendment
• Two-ceiling distinction: explicit and honest — "this seat's Electron binary aborted with SIGABRT" is a named sandbox ceiling, not an unprobed gap
• Evidence-class collapse check: pass — the body never promotes L2 to L3 framing
• Deployment causality: N/A — no external deploy gate

Reviewer-supplied L3, this host (Electron-capable), exact head 54b5cfefb0: AC4 receipt — npm --prefix harness run smoke:brain: passed: true, brainUp: true (the isolated organism came up: orchestrator marker + allocated-port Chroma + a real fleet wire round-trip, then clean group teardown), popup + shared-heap evidence ✓, rendererErrors: [], both plane leaves blanked as the new witnesses assert. One EXTRA finding this receipt surfaces, explicitly OUTSIDE this PR's scope: with the brain leg ready BEFORE window observation, the verdict still records rosterState: 'sample' — the cockpit renderer never went live against a ready owned transport; that is the #16699-family renderer-load defect (upgraded and claimed by this reviewer), orthogonal to the classifier under review. AC2's live fail-closed witness was deliberately NOT run in this review window (the operator flagged desktop window noise from verification runs mid-review; the witness opens another headed instance) — the down-plane property is structurally covered by the probe-free classifier witnesses and the residual stays where the author's list put it, unweakened.

Findings: Pass — achieved evidence exceeds the PR's declared ceiling on AC4; residual narrows to the operator-run items.

🧪 Test-Evidence & Location Audit

• Execution evidence: exact-head required CI green at 54b5cfefb0 (verified pre-review); author per-surface receipt: 117/117 harness tree at head — independently reproduced on this host: 117/117
• Reviewer falsifier: the AC4 smoke receipt — command npm --prefix harness run smoke:brain, named concern "isolated organism viability on an Electron-capable host", result passed (brainUp: true, clean teardown)
• Test location: classifier witnesses land in test/playwright/unit/harness/brain.spec.mjs beside the sibling brain.mjs exports — canonical

Findings: Pass.

N/A Audits — 📡 🔗

N/A across listed dimensions: no OpenAPI surface touched; no new skill/convention/MCP tool — the classifier + runbook consume established patterns.

📋 Required Actions

No required actions — eligible for human merge.

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 97 - the probe-free pure classifier beside its consumers, leaf resolution through the established config-resolver child, and the smoke containment strengthening are all exactly-placed; 3 held back for the failure-path mode-observability gap (the plan that selected a refused boot is invisible in the log line — a diagnosis affordance, not a placement fault).
• [CONTENT_COMPLETENESS]: 98 - JSDoc on the classifier names the down-plane rationale; README truth table + module-comment rewrite keep every prose surface coherent with the new order; 2 held for the same failure-path line not documenting the plan it refused under.
• [EXECUTION_QUALITY]: 98 - checked and cleared: the trim/whitespace edge (witnessed), plane+orchestratorAlive precedence (witnessed), foreign-listener still-throws in brain mode (diff-verified at the fleetPortHeld guard), packaged-product profile untouched while packaged-SMOKE blanks the leaves (both hunks verified), and the reviewer-run smoke receipt.
• [PRODUCTIVITY]: 100 - ACs 1/3/5 delivered; AC4's declared residual was closed by a reviewer-supplied receipt; the smoke-leak closure is bonus scope that belongs.
• [IMPACT]: 80 - this is the safety gate for every future start:brain on every dockerized-canonical machine; it converts today's live-fired hazard (own-mode beside a plane) from possible to structurally unreachable.
• [COMPLEXITY]: 40 - one pure function, two consumption points, two doc surfaces, profile blanks; the cognitive load lives in WHY (the ordering argument), which the prose now carries.
• [EFFORT_PROFILE]: Quick Win - high safety ROI over a small, witnessed surface.

The ticket said "the app must not ask the operator to do this manually"; this PR's version is better: the app no longer asks the QUESTION whose wrong answer was dangerous. Merge-eligible.

— Clio (@neo-fable-clio, Claude Fable 5, Claude Code) 📜