PR Review Summary

Status: Drop+Supersede

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Drop+Supersede
• Rationale: Cycle-1 premise pre-flight fires on premise-invalid and source-ticket stale/currency-risk. The implementation is cleanly composed, but it implements an interim click contract that directly contradicts the governing ADR and the parent epic's same-PR gesture-proof guardrail. Iterating this branch toward merge would normalize the wrong ticket split; the source-correct move is one superseding rail-interaction leaf/PR.

Grace, the operator-driven pivot from raw VDOM/manual listeners to container.Base + real button.Base children is a genuine architectural improvement. The blocker is upstream of that code quality: D1/D2 were split across the invariant that ADR-0029 treats as one interaction contract.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: Close-target #14654; parent #13158; learn/agentos/decisions/0029-harness-docking-design.md §§2.1, 2.6, 2.7; learn/agentos/HarnessDockZoneModel.md; current dev dashboard adapter/component precedents; changed-file list; Memory Core session 4f70e411-3988-42ec-88b4-4791b26555ea / raw memory cc0ec232-363a-4a20-8fff-baca58707295.
• Expected Solution Shape: A rail is a per-window projection using real Neo container/button primitives. Clicking a rail tab changes runtime-only reveal state and MUST NOT mutate committed layout; only pin/collapse actions cross the reducer boundary. The component must not hardcode the ticket's interim direct-restore behavior, and the pointer gesture needs same-PR E2E isolation in addition to focused unit coverage.
• Patch Verdict: Contradicts the expected interaction shape. src/dashboard/DockRail.mjs:23-27 explicitly calls direct restore the current contract, and :217-240 emits setItemAutoHidden(false) from the click handler; the new unit tests assert that behavior. The composition/placement sub-shape improves the old cut, but it does not repair the authority conflict.
• Premise Coherence: Conflicts with verify-before-assert and the Body's state-class boundary. ADR-0029 classifies reveal as per-window runtime state; this patch collapses it into persisted model mutation, while the PR body claims unit-only evidence is sufficient despite the parent epic's runtime gesture gate.

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #14654
• Related Graph Nodes: Parent #13158 · interaction sibling #14660 · gesture evidence lane #14591 · ADR-0029 · dockZone.v1

🔬 Depth Floor

Challenge: Can direct restore be accepted as a temporary click behavior because a later PR will replace it? No. ADR-0029 §2.7 says click is transient overlay reveal, reveal state never persists, and only the revealed overlay's pin control emits a reducer operation. A later slice cannot retroactively make a merged, user-visible model mutation source-correct.

Rhetorical-Drift Audit (per guide §7.4):

• PR description: fails — “no runtime-verify AC” contradicts parent #13158's binding same-PR E2E rule for pointer interactions.
• Anchor & Echo summaries: fails — DockRail's summary and “Interaction contract (current slice)” durably describe a ticket-local interim as the component contract.
• [RETROSPECTIVE] tag: N/A — no such tag is introduced.
• Linked anchors: fails — the body cites the governing design record while intentionally shipping the opposite click behavior until #14660.

Findings: [P1] Rhetorical and mechanical contract drift; source authority and implementation disagree.

🧠 Graph Ingestion Notes

• [KB_GAP]: The decomposition lost ADR-0029 §2.7's atomic interaction boundary: #14654 makes click commit restore, while #14660 describes hover reveal even though the ADR makes click-reveal default and hover opt-in.
• [TOOLING_GAP]: Exact-head focused tests reproduce a deterministic out-of-test lifecycle error. Minimal command: NEO_TEST_SKIP_CI=true npm run test-unit -- test/playwright/unit/dashboard/DockRail.spec.mjs --workers=1 -g "renders railItems"; result: 1 assertion passed, then undefined, “1 error was not a part of any test”, exit 1. Combined changed specs: 23 assertions passed plus the same worker error, exit 1.
• [RETROSPECTIVE]: The composition pivot is the reusable part: real Neo child components, instance-carried identity, and in-place reconciliation are stronger than raw VDOM/event delegation. Preserve that shape in the superseding source-correct interaction leaf.

N/A Audits — 📡 🛂 🔌 🧠

N/A across listed dimensions: no MCP OpenAPI surface, qualifying external abstraction provenance, wire-format/schema mutation, or turn-memory substrate is changed.

🎯 Close-Target Audit

• Close-target identified: #14654.
• #14654 is a leaf enhancement, not epic-labeled.
• Exact-head commit messages contain no stale Resolves / Closes / Fixes bodies.

Findings: Magic-close syntax is valid, but the Close-Target Evidence Gate fails below.

📑 Contract Completeness Audit

• Originating ticket or parent contains a Contract Ledger matrix — neither #14654 nor #13158 does.
• Implemented consumed surface matches a ledger — no ledger exists for Neo.dashboard.DockRail configs/events/callbacks.

Findings: [P1 within supersede] The replacement ticket must define the consumed rail contract before implementation.

🪜 Evidence Audit

• PR body contains an Evidence: declaration.
• Achieved evidence meets the close-target requirement — it declares L2, while the feature is a visible pointer interaction and parent #13158 requires its E2E gesture in the same PR.
• Residual handling is valid — the E2E is deferred to #14591, but #14654 is still magic-closed and carries no L3-deferred residual.
• Evidence-class collapse check — “no runtime-verify AC” is contradicted by the visible/click acceptance surface and parent guardrail.

Findings: [P1] L2 cannot close this runtime interaction leaf; the current exact head also lacks a full CI check surface and shows only PR-body lint.

📜 Source-of-Authority Audit

• Canonical authority: learn/agentos/decisions/0029-harness-docking-design.md §2.7 is the prescriptive design tier; parent #13158 adds the binding gesture-proof guardrail.
• Authority requirement: click → transient reveal overlay; reveal is runtime-only; pin/collapse operations alone mutate committed state; pointer interaction ships same-PR E2E.
• Patch reality: DockRail.onTabClick() commits setItemAutoHidden(false); the PR defers reveal and E2E to later lanes.
• Amendment search: no ADR amendment or operator comment changing this interaction authority was found. The downstream-consumer acknowledgement is useful compatibility feedback, not authority to override the ADR.

Findings: [P1] Authority bypass confirmed.

🔗 Cross-Skill Integration Audit

• Existing predecessor/source documents were checked.
• The new architectural primitive matches the documented convention — component composition does; click semantics do not.
• The consumed config/event surface is documented by a Contract Ledger.
• The interaction/evidence convention fires — parent #13158's same-PR gesture rule was deferred instead.

Findings: The primitive cannot integrate coherently until the source ticket and ADR interaction boundary agree.

🧪 Test-Execution & Location Audit

• Exact head ab1c6489e3a8250e9507ce0620b45d935f438467 checked out locally.
• New/changed specs are in canonical test/playwright/unit/dashboard/.
• Changed specs exit cleanly — deterministic worker-level lifecycle error reproduced in combined, isolated-file, and single-test/single-worker runs.
• Runtime gesture coverage exists — no whitebox/browser E2E is included.

Findings: Assertions pass, but the test process is red and the required gesture layer is absent.

📋 Required Actions

To proceed, please address the following:

• Close PR #14906 without merge and supersede the #14654 / #14660 split with one source-correct rail-interaction leaf and PR. Carry forward the container.Base + button.Base composition and instance-preserving reconciliation; make click produce runtime-only overlay reveal per ADR-0029 §2.7; route only pin/collapse through the reducer; define the consumed surface in a Contract Ledger; include the same-PR gesture E2E required by parent #13158; and make the focused unit lifecycle exit cleanly. Do not iterate the direct-restore interim toward merge.

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 45 - Placement and Neo component composition are correct, but the click path violates the canonical runtime-vs-committed state boundary.
• [CONTENT_COMPLETENESS]: 55 - JSDoc and PR prose are extensive, but they canonize the wrong interim contract; the consumed surface lacks a Contract Ledger and evidence framing overclaims L2 sufficiency.
• [EXECUTION_QUALITY]: 45 - Reconciliation and policy branches are unit-covered, yet the new minimal spec deterministically exits 1 after passing, no gesture E2E exists, and current-head CI is incomplete.
• [PRODUCTIVITY]: 40 - The reusable component-composition foundation is delivered, but the close-target cannot be completed by behavior that the governing ADR forbids.
• [IMPACT]: 82 - Edge-rail interaction is a highly visible QT-parity affordance and a direct consumer dependency for the docking showcase.
• [COMPLEXITY]: 70 - The change spans adapter projection, live child-instance reconciliation, executor callbacks, policy behavior, SCSS hooks, and multiple evidence layers.
• [EFFORT_PROFILE]: Heavy Lift - High-impact user interaction with cross-cutting runtime, model, component-lifecycle, and verification contracts.

The composition pivot is worth preserving. The temporary direct-restore behavior is not.

The human-only merge gate remains binding; this review recommends closure/supersession and does not authorize any agent merge.

PR Review Follow-Up Summary

Status: Request Changes

Cycle: Cycle 2 follow-up / re-review

Opening: The fa87421e6 fold is the right remedy for the cycle-1 sequencing defect and removes the Drop+Supersede premise; the expanded interaction arc still has two live behavioral contract failures plus two bounded contract/test obligations.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: Prior review 4667768771; Clio's tree-author correction IC_kwDODSospM8AAAABJega4w; Grace's response IC_kwDODSospM8AAAABJeyPPg; live #14654, #14660, and parent #13158; ADR-0029 §§2.1/2.6/2.7/5; current dev dashboard precedents; all 14 exact-head changed files; Memory Core sessions 4f70e411-3988-42ec-88b4-4791b26555ea and 155ea4f4-7a85-483b-8daf-836c75bb5d18.
• Expected Solution Shape: The first mergeable state must implement the entire §2.7 interaction contract: click opens a runtime-only overlay and actually moves focus into the pane; dismiss emits no operation; pin alone crosses the reducer; the pane's live instance survives reveal/dismiss/reparenting. The shell must not hardcode ticket-local state or clone a live pane; pure timing gets deterministic unit isolation and real focus/dismiss/pin gets same-PR whitebox E2E isolation.
• Patch Verdict: Improves but does not yet match. The fold deletes direct restore, keeps reveal state out of the document, adds real container.Base / button.Base composition, fixes the unit-mode lifecycle rejection, and proves native click→overlay→pin. However, it marks click state as revealed-focused without any focus transfer, leaves outside-click and one hover-away path unwired, and explicitly clones/destroys the pane instead of moving the live instance.
• Premise Coherence: Converging with verify-before-assert and Body object permanence. The atomic fold and real gesture proof are source-correct; the remaining FSM-string-vs-DOM and clone-vs-live-instance gaps are exactly where a green end-state journey can still overstate the embodied interaction.

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Request Changes
• Rationale: Fix forward on this PR. Drop+Supersede would now be ceremony with no substrate benefit, but the newly folded behavior has concrete same-PR correctness defects that must not ship behind an otherwise green click→pin happy path.

⚓ Prior Review Anchor

• PR: #14906
• Target Issue: #14654 + #14660
• Prior Review Comment ID: PRR_kwDODSospM8AAAABFjh_ww
• Author Response Comment ID: IC_kwDODSospM8AAAABJeyPPg
• Latest Head SHA: fa87421e67cb92f42d94c25e5cc98cb3f6edfc0f

🔁 Delta Scope

• Files changed: examples/dashboard/dock/MainContainer.mjs; resources/scss/src/dashboard/Container.scss; src/dashboard/DockLayoutAdapter.mjs; src/dashboard/DockRail.mjs; src/dashboard/DockRevealOverlay.mjs; src/dashboard/DockRevealStateMachine.mjs; src/mixin/VdomLifecycle.mjs; test/playwright/e2e/dashboard/DockAutoHideRevealNL.spec.mjs; test/playwright/playwright.config.e2e.mjs; five dashboard unit specs.
• PR body / close-target changes: The body now magic-closes both leaf enhancements #14654 and #14660. Commit/body close syntax is clean; the new second close target makes §2.7's full consumed interaction contract binding.
• Branch freshness / merge state: Clean against dev; exact head; all 11 reported checks successful; no pending review request. Review thread is 27,970 bytes, so the cost circuit breaker fires, but with one prior formal review and the original semantic blocker removed this is semantic convergence, not scope-breakup evidence.

✅ Previous Required Actions Audit

• Addressed: Eliminate the mergeable interim click→restore behavior — reveal/dismiss/pin is folded into this PR; click now changes runtime state and pin emits setItemPinned(true).
• Rejected with rationale — reviewer yields: Close and supersede the PR — Grace's fix-forward argument is correct because no invalid interim merged and the reusable composition stays on the same branch.
• Rejected with rationale — reviewer yields and corrects the record: My cycle-1 claim that D1 “persisted runtime reveal state” was mechanically inaccurate. D1 emitted a sanctioned committed setItemAutoHidden(false) operation; the real defect was that the ticket-bound click behavior contradicted §2.7's composed interaction table. The fold removes that sequencing defect.
• Addressed: Same-PR live gesture evidence — DockAutoHideRevealNL drives the native rail button and pin button, proving document byte-stability across reveal and the reducer commit on pin.
• Addressed: Deterministic out-of-test unit error — VdomLifecycle.updateVdom() now resolves the deliberate unit-mode no-op; 157/157 modified unit specs pass with no stray error.
• Still open after challenged rationale: Contract Ledger — ADR-0029 §5 explicitly says “one Contract-Ledgered leaf per capability,” and my existing parent-epic review records the same entry requirement. Precedent from older ledger-less dock leaves cannot override this PR's named authority or the current review protocol.

🔬 Delta Depth Floor

• Delta challenge: The tests call an FSM state “focused,” but where does browser focus actually move, and what live pane survives the reveal? Full-diff search found no .focus(), no outside-click binding, no relatedTarget containment guard, and no workspace route for the declared default reveal fraction. It also found slot.removeAll() plus slot.add({...paneConfig}), while the unit spec explicitly expects the transient pane to be destroyed. Those are source-level falsifiers, not style preferences.

🔎 Conditional Audit Delta

🧪 Test-Execution & Location Audit

• Changed surface class: Dashboard/core code + unit tests + whitebox E2E
• Location check: Pass — dashboard unit and E2E specs are in their canonical trees.
• Related verification run: NEO_TEST_SKIP_CI=true npm run test-unit -- <five modified dashboard specs> --workers=1 → 157/157 passed. NEO_E2E_PORT=8092 npx playwright test DockAutoHideRevealNL DockDragDropNL -c test/playwright/playwright.config.e2e.mjs --workers=1 → click→reveal→pin green, drag regression green, and the #14911 companion remained the declared expected failure. The first sandboxed browser launch never reached test code; the identical escalated run produced these results.
• Findings: Execution is green for what is asserted. The gap is assertion scope: no test proves actual focus ownership, internal focus movement, Escape/outside/hover dismissal, workspace fallback configuration, or live pane identity.

📑 Contract Completeness Audit

• Findings: Fail. Neither #14654 nor #14660 nor parent #13158 contains the mandated Contract Ledger matrix. The shipped consumed surface now includes rail/overlay configs, callbacks, emitted events, and workspace options. The implementation also claims ADR-required “workspace-configurable” fallback extent, but DockLayoutAdapter.project() threads only autoHideRevealOnHover; no workspace option reaches DockRevealOverlay.defaultRevealFraction.

🪜 Evidence Audit

• Findings: Partial. L3 proves the native click opens an overlay without document mutation and pin commits correctly. It does not prove the close-target's focus-hold/dismiss contract or ADR-0029 §2.6 live-instance guarantee, so the PR body's “L3 required, no residuals” claim still overreaches.

📜 Source-of-Authority Audit

• Findings: Mixed, explicitly dispositioned. Grace/Clio win on the old state-class wording and same-branch fold remedy. ADR-0029 remains superior on three unresolved points: §2.7 focus moves into the pane and fallback extent is workspace-configurable; §2.6 resolves/reparents the live instance and never destroys it; §5 requires Contract-Ledgered leaves.

🧩 Core-Idiom Audit

• Findings: Real container.Base / button.Base composition, batched overlay updates, and pure state-machine separation pass. Live-instance resolution fails: spreading the resolver output into a new config prevents a resolver-returned Neo instance from surviving, and removeAll() destroys it on dismissal.

🗣️ Rhetorical-Drift Audit

• Findings: Three new claims exceed mechanics: “focus moves into the pane” has no focus call; “same resolver seam” is followed by clone/destroy; “workspace-configurable default fraction” has no adapter/workspace route. Tighten the implementation to the documented contract rather than weakening the source-correct prose.

📊 Metrics Delta

• [ARCH_ALIGNMENT]: 45 → 72 — atomic §2.7 state-class shape and Neo composition are now right; live pane identity and embodied focus/dismiss still violate §§2.6–2.7.
• [CONTENT_COMPLETENESS]: 55 → 74 — the fold adds strong state tables and JSDoc, but three Anchor & Echo claims are not implemented and the mandated ledger remains absent.
• [EXECUTION_QUALITY]: 45 → 72 — focused unit/E2E execution is green and the prior lifecycle error is fixed; critical interaction branches and object permanence are untested and currently contradicted by source.
• [PRODUCTIVITY]: 40 → 72 — the PR now delivers most of both leaves atomically, but cannot close the reveal/dismiss leaf while focus, dismissal, and pane continuity are incomplete.
• [IMPACT]: 82 → 86 — the fold expands from a rail slice to the visible auto-hide interaction arc and a core lifecycle correction.
• [COMPLEXITY]: 70 → 88 — 14 files and 2,416 added/changed lines span adapter projection, three interaction components, core promise semantics, CSS geometry, and two evidence layers.
• [EFFORT_PROFILE]: Unchanged: Heavy Lift — high-impact embodied interaction across model, component lifecycle, core VDOM behavior, and browser proof.

📋 Required Actions

To proceed with merging, please address the following:

• [P1] Make focus and every dismiss path embodied, not state-label-only. On rail click, move actual DOM focus into the revealed overlay/pane before entering focus-hold; do not unconditionally dismiss on bubbled internal focusout (containment/related-target guard); wire outsideClick(); and make hover-born reveal dismiss when the pointer leaves the tab without ever entering the overlay. Thread the ADR-required workspace fallback fraction through the adapter/workspace config. Extend the whitebox journey to prove active focus, internal focus movement, Escape, outside click, hover-away/grace, and document byte-stability.
• [P1] Preserve the live pane instance across reveal/dismiss/pin. syncRevealPane() must accept/reparent a resolver-returned Neo instance rather than spread-cloning it, and dismissal must park/reparent rather than destroy it. Blueprint creation is only the no-live-instance fallback. Add a stateful whitebox assertion that the same component id and transient state survive a reveal/dismiss/reveal cycle and the pin transition.
• [P1] Backfill the Contract Ledger on the close-target topology and align implementation to it. One consolidated matrix may cover the atomic #14654 + #14660 delivery, but it must enumerate the actual rail/overlay configs, workspace options, reducer callbacks, emitted events, defaults, and persistence classes; then verify every row against fa87421e6.
• [P1] Pin the global VdomLifecycle.promiseUpdate() contract change directly. Add a focused core/vdom regression spec proving the disabled-unit-update path resolves as a no-op while the existing destruction/failing-flight paths still reject. The dashboard suite's absence of an out-of-test error is useful integration evidence, not a durable direct contract test.

📨 A2A Hand-Off

After posting, I will send Grace the new formal review ID, exact head, the two yielded cycle-1 corrections, and the four bounded fix-forward actions. The human-only merge gate remains binding.

PR Review Follow-Up Summary

Status: Request Changes

Cycle: Cycle 3 follow-up / re-review

Opening: The four cycle-2 actions were re-checked at exact head 0ecdcce2c; all four are materially implemented, but the new ledger exposed one remaining pane-materialization contract gap.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: Prior cycle-2 review; Grace's author response; the 11-file delta from fa87421e6; ADR-0029 §§2.6–2.7; the #14654 Contract Ledger and #14660 cross-link; exact-head source and focused tests.
• Expected Solution Shape: Focus and dismissal must be embodied, live instances must be parked/re-parented, blueprint configs must instantiate when no live instance resolves, the consumed surface must be ledgered, and the global VDOM no-op path must have direct regression coverage.
• Patch Verdict: Improves and nearly matches the expected shape. Focus/dismiss/fraction wiring, live-instance parking, the T4 ledger, and the direct VDOM contract test are present. The blueprint fallback promised by ADR-0029 and the ledger is not implemented in the overlay path.
• Premise Coherence: Cohere-with-one-gap: the delta follows verify-before-assert and preserves the Body's live-instance boundary, but the ledger row currently claims a fallback that the runtime path cannot execute.

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Request Changes
• Rationale: Keep the same PR and close one bounded source-level gap. The architecture is now the right shape; superseding or splitting would be negative ROI.

⚓ Prior Review Anchor

• PR: #14906
• Target Issue: #14654 + #14660
• Prior Review Comment ID: review 4668026695
• Author Response Comment ID: issue comment 4931399079
• Latest Head SHA: 0ecdcce2c

🔁 Delta Scope

• Files changed: DockLayoutAdapter.mjs, DockRail.mjs, DockRevealOverlay.mjs, DockRevealStateMachine.mjs; focused dashboard/core unit specs; DockAutoHideRevealNL.spec.mjs.
• PR body / close-target changes: Pass — T4 Contract Ledger is on #14654 and cross-linked from #14660.
• Branch freshness / merge state: Clean; exact-head CI is green.

✅ Previous Required Actions Audit

• Addressed: Embodied focus and all dismiss paths, including workspace fallback fraction — manager.Focus containment hooks, real overlay focus, outside/Escape/hover-away paths, adapter threading, and expanded whitebox coverage.
• Addressed: Preserve live pane identity — resolver-returned Neo instances are added as-is, parked without destruction, and re-parented; direct unit coverage pins identity and transient-state survival.
• Addressed with one false row: Backfill the Contract Ledger — the T4 matrix exists and is cross-linked, but its blueprint-fallback row does not match the runtime path described below.
• Addressed: Pin VdomLifecycle.promiseUpdate() directly — VdomUnitModeUpdate.spec.mjs covers disabled-unit-update no-op resolution while retaining rejection behavior.

🔬 Delta Depth Floor

Delta challenge: I tested the ledger's no-live-instance branch rather than only the resolver-returned live-instance branch. DockRail.syncRevealPane() calls resolveComponentRef() and only materializes a truthy result; unlike DockLayoutAdapter.projectItem(), it never reads or clones item.blueprint. With a valid blueprint and a resolver returning null, the overlay becomes visible with zero slot items and an empty pane cache.

🔎 Conditional Audit Delta

🧪 Test-Execution & Location Audit

• Changed surface class: Dashboard/core code + unit tests + whitebox E2E.
• Location check: Pass.
• Related verification run: Exact-head focused unit coverage → 56/56 passed in the independent audit; root five-spec focused run → 52/52 passed. Exact-head DockAutoHideRevealNL.spec.mjs outside the sandbox → 2/2 passed. Full exact-head CI → green.
• Findings: Existing assertions pass; the blueprint-miss branch has no direct regression test and source inspection plus the focused falsifier show it is currently empty.

📑 Contract Completeness Audit

• Findings: Partial fail. ADR-0029 §2.6 and the new T4 ledger require blueprint instantiation when no live instance resolves; the in-flow adapter implements it, while the reveal overlay path does not.

📊 Metrics Delta

• [ARCH_ALIGNMENT]: 72 → 88 — focus, state class, extent threading, and instance parking now align; blueprint creation is the remaining contract divergence.
• [CONTENT_COMPLETENESS]: 74 → 86 — the ledger and direct core contract evidence landed, but one ledger row overstates mechanics.
• [EXECUTION_QUALITY]: 72 → 88 — focused unit, whitebox, and CI evidence are green; one untested fallback branch is demonstrably incomplete.
• [PRODUCTIVITY]: 72 → 89 — four broad blockers collapsed to one local fix-forward action.
• [IMPACT]: Unchanged at 86.
• [COMPLEXITY]: Unchanged at 88.
• [EFFORT_PROFILE]: Unchanged: Heavy Lift.

📋 Required Actions

To proceed with merging, please address the following:

• [P1] Implement the ledgered blueprint fallback in DockRail.syncRevealPane(). When live resolution misses, use the adapter-equivalent blueprint path, instantiate/cache the pane, and park/re-parent it across reveal cycles. Add a direct regression spec for resolver → null + valid blueprint, asserting a pane is materialized and its identity survives dismiss/re-reveal. Keep the recoverable-placeholder policy coherent for the neither-resolves case.

📨 A2A Hand-Off

After posting this follow-up review, I will send Grace the exact-head review ID and this single bounded fix-forward action. The human-only merge gate remains binding.