# Phase 180 Gap Analysis — Remote Creative Control Plane

Critical review of `TODOS/phase-180.md` (`180.C.01` → `180.C.30`). For each gap,
the heading states the gap, the body states what the checklist has today, what
is missing, and what an ordered follow-up task would look like if added. Items
are organized by concern area, not by stage, because most gaps cut across
multiple stages.

Scope: this is a gap review, not a critique of checklist quality. The
sequencing, policy-first discipline, audit-before-mutation rule, and stub-ban
hygiene in `180.C.01`/`180.C.07`/`180.C.08` are strong and should be preserved.

---

## 1. Capability Surface — Platforms and Tools

### 1.1 macOS-only host ignores the working surface of most studios

- **Has:** Every host-side stage (`180.C.06`, `180.C.13`, `180.C.26`) assumes
  macOS: TCC permissions, Keychain, LaunchAgent, notarization, menu-bar app.
- **Missing:** Windows and Linux remote host. Blender and Unreal studios run
  predominantly on Windows workstations; render nodes usually run Linux. The
  phase title says "MacBook" but the broader vision ("Hetzner, AWS, or any other
  trusted compute environment") implies heterogeneous hosts.
- **Follow-up:** Add explicit stages for Windows host packaging (Services,
  code-signing, UAC, Defender), Linux host packaging (systemd user units,
  Wayland vs X11 screen capture, PipeWire), and a cross-platform
  permission/discovery abstraction in `mac-host-runtime`. Or scope the phase
  title down to macOS and defer other OSes to a future phase — but make the
  decision explicit.

### 1.2 DCC coverage stops at Blender and Unreal

- **Has:** Deep adapters for Blender (`180.C.09`/`.10`/`.17`/`.18`) and Unreal
  (`180.C.15`/`.23`/`.24`); generic browser and desktop fallback.
- **Missing:** Maya, Houdini, Cinema 4D, Nuke, DaVinci Resolve, Substance
  Painter/Designer, ZBrush, Photoshop/After Effects, Logic/Pro Tools/Ableton,
  Figma desktop, Procreate. Desktop fallback can reach these but no structured
  adapter path exists. The agent tool-selection policy in `180.C.14.10` will
  collapse to "desktop fallback" for every non-Blender/Unreal DCC.
- **Follow-up:** Add a generic Adapter SDK stage (see §5.4) so third parties and
  later phases can land Maya/Houdini/Substance/etc. without reopening the
  protocol. At minimum, name the "Tier 2 DCC" backlog explicitly so the
  Blender/Unreal rigor isn't misread as the full scope.

### 1.3 Audio/video/realtime pipelines are absent

- **Has:** Render artifacts (`180.C.18.05`, `180.C.24.03`), Movie Render Queue.
- **Missing:** Audio workflows (DAW transport control, timecode sync, Pro
  Tools/Logic session state), video editing (Resolve/Premiere timeline and
  project queries), Live Link / virtual production control beyond the read-only
  observation in `180.C.24.08`, OCIO/ACES color management across DCCs, HDR
  stream handling, frame-accurate scrubbing.
- **Follow-up:** Add a "Timecode, Color, and Media Pipeline" stage between
  `180.C.22` and `180.C.23`; define an `OCIOConfigReference` shared schema; and
  mark audio DCC support as in-scope or out-of-scope explicitly.

### 1.4 No mobile or tablet host/observer

- **Has:** Mobile observer/approval is a single bullet at `180.C.30.09` — "after
  the desktop Control Room is stable."
- **Missing:** Mobile doesn't need to wait for 30 stages to be useful —
  approving a deny/approve prompt from a phone is one of the highest-value
  operator moments. Voice ("Iris supervision") is buried in the same bullet.
- **Follow-up:** Pull approval-only mobile (PWA or native wrapper) into a
  `180.C.12.x` or `180.C.21.x` substage so operators aren't tethered to a laptop
  just to approve an asset import.

### 1.5 Stream capabilities are thin

- **Has:** WebRTC MVP (`180.C.12.07`), source switching (`180.C.21.03`),
  overlays (`180.C.21.04`), TURN fallback (`180.C.25.04`).
- **Missing:** HDR/wide-gamut stream, multi-display layout stream, hardware
  codec selection (VideoToolbox/NVENC), bandwidth-adaptive bitrate, audio stream
  (not just screen), cursor capture as separate track, region-of- interest
  encoding, stream recording with keyframe snapshot index for fast seeking,
  picture-in-picture on operator side, and zoom/pan on the viewport.
- **Follow-up:** Expand `180.C.12` or add a dedicated streaming stage that
  enumerates codec, color, audio, multi-track, recording, and scrubbing.

### 1.6 Asset provenance has gaps around dependencies

- **Has:** Artifact metadata, hashes, provenance links (`180.C.22.05`), cross-
  app transfer flows (`180.C.22.06`–`180.C.22.09`).
- **Missing:** Dependency graph for derived assets (if texture X is replaced,
  which Blender materials and Unreal material instances depend on it?), broken
  reference detection, upstream/downstream impact preview before edits.
- **Follow-up:** Add to `180.C.22` or a new stage: asset dependency index,
  reference integrity queries, impact-preview command used by `180.C.14.14`
  dry-runs.

---

## 2. UX — Control Room and Operator Surface

### 2.1 Operator onboarding is unspecified

- **Has:** Host first-run onboarding is thorough (`180.C.26.03`).
- **Missing:** Operator-side first-run UX in the Control Room: empty states
  before any device is paired, guided pairing from the operator side,
  explanations of what "observe-only vs project write" means in plain language,
  and the "I just opened this; what do I do?" flow.
- **Follow-up:** Add to `180.C.12` (MVP) and expand in `180.C.21`: empty states,
  pairing wizard, permission-meaning tooltips, and a "first session" guided
  tour.

### 2.2 Notification and alert channels for operators are undefined

- **Has:** In-app timeline, dashboards, alerts (`180.C.28.02`).
- **Missing:** Operator reachability — desktop notifications, email, Slack, SMS,
  or mobile push for approval-required, host-offline, long-running completion,
  and policy denials. `180.C.28.02` is for ops dashboards, not the creative
  operator.
- **Follow-up:** Add a notifications stage or items in `180.C.21`: channel
  preferences per user, per event class, with snooze/mute rules and test
  delivery.

### 2.3 Keyboard, accessibility, and localization are partial

- **Has:** Playwright + accessibility tests for Control Room states
  (`180.C.21.09`).
- **Missing:** Explicit keyboard shortcuts for approve/deny/pause/takeover,
  command palette, screen-reader-friendly timeline semantics, high-contrast and
  reduced-motion modes, localization (strings, dates, numbers, RTL),
  international keyboard layout handling for desktop-fallback typing.
- **Follow-up:** Add substages: keyboard-shortcut map and conflict check;
  i18n/l10n string extraction; RTL and reduced-motion themes; and a
  keyboard-layout-aware `desktop.type` command.

### 2.4 Stream scrubbing, annotation, and sharing

- **Has:** Audit replay view (`180.C.21.07`), replay export (`180.C.28.04`).
- **Missing:** Frame-accurate scrubbing with keyframe index, in-timeline
  annotations by the operator ("this is where the agent went wrong"), named
  bookmarks, clip/moment sharing with signed expiring links (for post-incident
  reviews), diff view for file/scene changes visualized on stream segments.
- **Follow-up:** Split `180.C.21.07` into discrete items: keyframe index,
  annotation layer, bookmark persistence, expiring share link with audit.

### 2.5 Cost visibility to the operator is missing

- **Has:** Spend and usage controls (`180.C.27.07`).
- **Missing:** Live, per-session cost display — tokens, gateway compute, TURN
  bandwidth, render/build minutes — in the Control Room while it happens, not
  just retroactively in budgets. Operators should be able to see "this session
  has cost $0.47 so far" before approving the next render.
- **Follow-up:** Add "Live cost panel" to `180.C.21` with per-command cost
  attribution and session running total.

### 2.6 Multi-session awareness

- **Has:** Single-session page (`180.C.21.02`).
- **Missing:** Multi-session overview — when an operator supervises three agents
  in parallel, they need a grid/mosaic view of mini-streams, a consolidated
  approval queue across sessions, and a "cross-session risk" ranking so they
  know which session to focus on.
- **Follow-up:** Add a multi-session dashboard item to `180.C.21` or in
  `180.C.30.03` (multi-host rooms).

---

## 3. Ergonomics — Agent-Facing

### 3.1 Capability discovery ergonomics

- **Has:** `device.list`, capability reports (`180.C.06.04`, `180.C.14.03`),
  capability memory (`180.C.14.17`).
- **Missing:** A structured "what can I do right now, given current state"
  query. `RemoteCapability` lists what adapters exist; it does not tell the
  agent that Blender is currently modal in a popup, Unreal has 40 shaders
  compiling, or the project root is read-only today. Agents will burn tokens
  rediscovering state.
- **Follow-up:** Add a `state.snapshot` or `agent.readiness` command that
  returns current-state booleans and blockers alongside static capability.

### 3.2 Dry-run impact preview is only a single bullet

- **Has:** `180.C.14.14` — "dry-run plan and human-readable impact summary."
- **Missing:** This one bullet carries enormous weight. It needs: schema for the
  plan object, integration with every risky namespace (not just Blender/
  Unreal/browser but file/shell/desktop), an explicit cost/time estimate on top
  of impact, and rendering in Control Room so operators can compare "predicted
  vs actual" after execution.
- **Follow-up:** Split into `DryRunPlan` schema, per-namespace planners, cost/
  time estimator, Control Room predicted-vs-actual view.

### 3.3 No record/replay sandbox for agent authors

- **Has:** Mock harness for gateway/host/adapters (`180.C.04`).
- **Missing:** A record-and-replay mode where an agent author can capture a real
  session and replay it offline to iterate on planning heuristics
  deterministically. The mock harness supports writing tests; it does not
  support "run my agent against a known scene/page without a real Blender or
  Chrome."
- **Follow-up:** Add a simulation/replay mode to the mock harness that loads
  real captured snapshots and replays adapter responses.

### 3.4 Agent error ergonomics

- **Has:** `RemoteErrorCode`, `RemoteWarning` (`180.C.03.05`), failure taxonomy
  (`180.C.28.06`).
- **Missing:** Guidance on _what the agent should try next_ embedded in the
  error. A `permission.missing` error with "go ask the operator to grant Screen
  Recording and retry this command" is dramatically more useful than a bare
  code. Structured `remediation` hints with a machine-parseable
  suggested-next-command would reduce wasted LLM turns significantly.
- **Follow-up:** Extend `RemoteErrorCode` with structured `remediation` field:
  human text + suggested command template + whether auto-retry is safe.

---

## 4. Ergonomics — Developer and SDK

### 4.1 Local dev quickstart isn't staged

- **Has:** Gateway docs (`180.C.05.09`), host docs (`180.C.06.09`), operator
  docs (`180.C.29.02`).
- **Missing:** A single "5-minute from clone to round-trip" quickstart —
  docker-compose for gateway+postgres+redis, a fake host bound to localhost, an
  example MCP client, and a working Blender mock adapter. Each piece exists in a
  different stage's docs; new contributors have to reconstruct the spine.
- **Follow-up:** Add a `180.C.02.x` or `180.C.05.x` item for a single top-level
  quickstart referencing the individual package docs.

### 4.2 Adapter SDK for third parties

- **Has:** `libs/bellona/remote-adapters` adapter interfaces (phase-180
  introduction).
- **Missing:** A public, documented adapter SDK with stable contract, hot-
  reload for dev, conformance test suite that a Maya/Houdini/Substance plugin
  author could target without forking Bellona. Without this, §1.2 can never be
  addressed cleanly.
- **Follow-up:** Add an "Adapter SDK and Conformance Suite" stage after
  `180.C.15` (once Blender + Unreal have validated the adapter interface by
  example).

### 4.3 Command-author workflow is undefined

- **Has:** Developer docs for adding commands, adapters, MCP tools, policy
  rules, etc. (`180.C.29.03`).
- **Missing:** A scaffold or generator. Right now adding `foo.bar` means
  touching protocol, gateway, host, policy, audit, Control Room, SDK, CLI, docs
  — all by hand. A scaffolding tool that emits correct stubs across all of those
  is the difference between a 2-day task and a 2-hour task.
- **Follow-up:** Add `tools/bellona/new-command` generator with conformance test
  as part of `180.C.29.03`.

### 4.4 Protocol evolution beyond version negotiation

- **Has:** Protocol version negotiation and schema drift tests (`180.C.03.09`,
  `180.C.03.10`).
- **Missing:** Migration paths for stored data — device records, audit entries,
  session snapshots, approval decisions. What happens when the protocol adds a
  required field and a 6-month-old audit needs to be read?
- **Follow-up:** Add a stored-schema migration framework and tests to `180.C.08`
  (audit) and `180.C.07` (device/session storage).

---

## 5. Security, Privacy, and Trust

### 5.1 End-to-end encryption model isn't called out

- **Has:** TLS transport (`180.C.25.03`), device key rotation, pinning
  (`180.C.27.01`), project-root enforcement, redaction (`180.C.27.03`–`.05`).
- **Missing:** Is the gateway trusted-to-see-plaintext or zero-knowledge? If an
  organization uses a shared SaaS gateway, can they insist stream/content be
  E2E-encrypted between agent and host such that the gateway only routes? This
  is a foundational decision, not a late hardening one.
- **Follow-up:** Add an ADR in `180.C.02` deciding trusted-gateway vs E2E modes,
  and if E2E is adopted, corresponding key-exchange and audit integrity items.

### 5.2 Consent and recording law

- **Has:** Privacy/legal review (`180.C.27.10`).
- **Missing:** Explicit consent flows for subjects captured incidentally
  (FaceTime windows, colleagues on camera via screen share, dictation overlays),
  watermarking of recorded streams, retention defaults per jurisdiction,
  right-to-erasure workflows against audit trails.
- **Follow-up:** Break `180.C.27.10` into specific items: consent capture,
  watermarking, retention defaults, erasure-vs-immutable-audit reconciliation.

### 5.3 Supply-chain integrity of adapters

- **Has:** SBOM, vulnerability scanning, malware scanning, tamper detection
  (`180.C.26.10`–`.12`).
- **Missing:** Blender addons and Unreal plugins install from user-writable
  paths. A malicious Python addon silently added to Blender's addons directory
  would run inside the Remote Host's trust boundary. `180.C.26.12` checks host
  binaries; it should also continuously verify Blender addon/Unreal plugin
  hashes across sessions, not only at install.
- **Follow-up:** Add continuous integrity check for third-party-extended DCC
  trust boundaries to `180.C.26` and `180.C.27`.

### 5.4 Agent-to-agent trust and identity

- **Has:** Device identity, revocation, organization controls (`180.C.27.01`,
  `180.C.27.06`).
- **Missing:** Agent identity is conflated with device or session. When multiple
  agents (cloud, local, specialist) operate under one user (`180.C.14.16`),
  which agent took which action? Audit attribution should distinguish "human
  operator U → orchestrator agent A → Blender-specialist agent B → command" at
  every timeline entry.
- **Follow-up:** Formalize `RemoteActor` with operator/agent/delegation chain
  and thread it through audit and timeline.

---

## 6. Reliability, Operations, and Release

### 6.1 Rollback semantics aren't uniform

- **Has:** Blender undo checkpoints (`180.C.10.02`, `180.C.17.02`), Unreal
  editor transactions (`180.C.15.07`), file snapshots (`180.C.22.04`),
  failed-mutation rollback in recovery flows (`180.C.28.08`).
- **Missing:** Browser and desktop fallback have no rollback story. "Undo" for a
  form submission is a different flow than an editor `Ctrl+Z`, and there's no
  compensating-action framework. A single `SessionRollback` envelope that
  describes what reversal means per namespace is absent.
- **Follow-up:** Define a `RollbackStrategy` enum (undo stack, snapshot restore,
  compensating action, manual remediation-only) per command class and enforce
  its presence in `RemoteCommandEnvelope`.

### 6.2 SLOs and error budgets aren't defined

- **Has:** Performance budgets (`180.C.16.06`, `180.C.24.09`), dashboards and
  alerts (`180.C.28.01`–`.02`).
- **Missing:** Explicit SLO targets (command acknowledgement, stream start,
  approval round-trip, audit write) and error-budget policy that gates release.
  Without SLOs, alerts become noise.
- **Follow-up:** Add SLO definition and error-budget policy to `180.C.16` or
  `180.C.28`, with automated release-gate check in `180.C.29`.

### 6.3 Release gate scope

- **Has:** Release gate item `180.C.16.09` lists the pre-stable checklist.
- **Missing:** It doesn't cover Unreal production (only Unreal MVP), browser
  production, desktop fallback hardening, macOS packaging, security completion,
  or fleet. The `180.C.29.08`/`.09`/`.10` final verification is more
  comprehensive but duplicates instead of superseding.
- **Follow-up:** Align `180.C.16.09` and `180.C.29.08`–`.10` — either the
  earlier gate is for "alpha" and clearly labeled, or it should point to the
  later one as the authoritative gate.

### 6.4 Migration from existing bridges

- **Has:** Audit of existing `bridge-blender`, `bridge-unreal`,
  `blender- agent`, `bridge-core` (`180.C.02.01`–`.04`).
- **Missing:** A deprecation/coexistence plan. What happens to consumers of the
  old bridges while the new Remote Host is rolled out? Is there a translation
  shim, a hard cutover, or parallel operation? Currently implicit.
- **Follow-up:** Add a coexistence/deprecation substage after `180.C.23` (Unreal
  production) describing shim, cutover date, and removal gate.

---

## 7. Testing

### 7.1 Fuzzing and adversarial schema tests

- **Has:** Extensive unit, integration, schema, contract, Playwright, and
  security tests across stages.
- **Missing:** Fuzz testing of `RemoteCommandEnvelope` against malformed inputs,
  oversized payloads, Unicode edge cases, and schema-downgrade attacks.
  Specifically called out because `180.C.27.09` focuses on prompt injection but
  not protocol-level fuzz.
- **Follow-up:** Add fuzzing to `180.C.04` (contract tests) and `180.C.27.09`
  (security tests).

### 7.2 Soak, load, and multi-day tests

- **Has:** Network impairment (`180.C.16.04`, `180.C.25.07`), chaos
  (`180.C.28.09`), performance budgets.
- **Missing:** Multi-day host uptime (sleep/wake cycles, LaunchAgent crash
  loops, slow memory leaks), gateway load tests (N hosts, M agents), audit store
  growth under sustained load. Absent from any stage.
- **Follow-up:** Add a reliability/soak substage to `180.C.25` or `180.C.28`.

### 7.3 Usability testing

- **Has:** Accessibility tests (`180.C.21.09`).
- **Missing:** Human usability evaluation of the approval flow, takeover flow,
  and multi-session supervision. "Can a new operator successfully approve the
  right command under time pressure?" is untested.
- **Follow-up:** Add user-research milestone to `180.C.21` (MVP) and `180.C.29`
  (pre-release).

---

## 8. Integration and Ecosystem

### 8.1 External event sinks

- **Has:** Internal audit, timeline, metrics.
- **Missing:** Outbound webhooks or event bus for Slack/Discord/email, ticketing
  (Linear/Jira), and incident systems (PagerDuty). Operators expect approval
  notifications and failure digests in their existing channels, not a separate
  Control Room.
- **Follow-up:** Add a webhooks/event-sink subsystem to `180.C.27` or
  `180.C.28`.

### 8.2 Production tracking integration

- **Has:** Cross-app creative flows (`180.C.22.06`–`.09`).
- **Missing:** Integration with production tracking (Autodesk Flow/Shotgun,
  ftrack, Kitsu). Studios will want agent actions reflected as task status
  updates automatically.
- **Follow-up:** Add optional production-tracking adapter to `180.C.22` or fleet
  stage.

### 8.3 MCP composition with other servers

- **Has:** MCP gateway (`180.C.14`), tool-selection policy (`180.C.14.10`).
- **Missing:** How Bellona MCP composes with other MCP servers (Isis, Yemaya,
  third-party). Agents in practice talk to multiple MCP servers; tool-name
  collisions, capability conflicts, and transport multiplexing are not
  addressed.
- **Follow-up:** Add an MCP composition substage to `180.C.14` covering
  namespacing, discovery, and priority conflicts.

---

## 9. Structural / Checklist Meta

### 9.1 Stage-size outliers

- `180.C.14` (MCP/Cloud-Agent) has 20 items; `180.C.26` has 14. Every other
  stage is ~10. Agents process "one stage" as one cognitive unit; oversized
  stages invite shortcut-taking, which `CLAUDE.md` already flags as a recurring
  failure mode.
- **Follow-up:** Split `180.C.14` into "MCP Gateway MVP" and "Cloud-Agent
  Operations / Agent Policy." Split `180.C.26` into "macOS Host App" and
  "Signing, Updates, SBOM."
- **Phase 180 disposition:** `180.C.14.26` records the Stage 14 split as a
  future planning rule without renumbering the executable checklist: `.01`-`.10`
  are MCP Gateway MVP and `.11`-`.25` are Cloud-Agent Operations and Policy. The
  `180.C.26` split remains open for `180.C.26.17`.

### 9.2 Cross-cutting concerns have no home stage

- **Has:** i18n, a11y, keyboard shortcuts, notifications, cost visibility,
  theming, localization — mentioned at most in passing.
- **Missing:** These concerns need to be staged alongside the features that
  introduce them, not left to "Control Room Production Experience" to retrofit.
  A "Cross-cutting UX Foundations" stage before `180.C.21` would catch them
  early.
- **Follow-up:** Add a cross-cutting UX stage or interleave the concerns as
  substages where they first become observable.

### 9.3 Hardware evidence budget

- **Has:** Hardware blocker rule (`180.C.01.04`), hardware smoke tests sprinkled
  across `180.C.09`–`180.C.24`.
- **Missing:** No explicit inventory — which MacBooks, which Blender versions,
  which Unreal versions, which Chrome channels — are required for sign-off.
  Without this, "hardware evidence" becomes unverifiable.
- **Follow-up:** Add a hardware-matrix item to `180.C.02` (alongside the
  compatibility matrix) listing required physical/virtual test rigs with OS/app
  versions.

### 9.4 "MVP" vs "Production" boundaries

- **Has:** Stage titles distinguish MVP (`180.C.09`, `.10`, `.11`, `.12`, `.13`,
  `.14`, `.15`) from Production (`180.C.17`, `.19`, `.21`, `.23`, `.24`).
- **Missing:** No single definition of what MVP ships with, what demo scenarios
  prove MVP-complete, or what the release artifact for MVP is. `180.C.16` is
  called "Production Hardening" but precedes the production-coverage stages.
- **Follow-up:** Rename `180.C.16` to "Alpha Hardening and MVP Release Gate" and
  add a list of MVP demo scenarios as sign-off criteria.

### 9.5 Unreal sequencing relative to phase title

- The phase title gives Blender and Unreal equal billing, but Unreal MVP is
  `180.C.15` (after Blender+browser+desktop+MCP) and production coverage is
  `180.C.23`–`.24`. If Unreal parity is a real goal, either accept the delay
  explicitly in the phase header or interleave Unreal MVP earlier.
- **Follow-up:** Add a note to the phase header setting Unreal expectations, or
  move Unreal MVP earlier.

---

## 10. Product / Go-to-Market

The checklist is an engineering plan. These items sit outside engineering but
block the system from being _used_ by anyone other than Oshun itself.

- **Pricing/metering model** — what usage is charged (seats, sessions, compute,
  bandwidth, render minutes)? `180.C.27.07` implements controls but not a model.
- **Customer segmentation** — solo creator vs small studio vs enterprise studio
  vs education. `180.C.27.11` references templates but not tiers.
- **Success metrics** — activation (first session), retention (weekly sessions),
  agent task completion rate, approval latency, incident rate. None currently
  tracked.
- **Free vs paid feature gating** — which MCP tools are open, which require
  enterprise plan.
- **Follow-up:** Product workstream parallel to `180.C.27`–`180.C.29`, or
  explicitly out of scope with a note.

---

## Summary of Top Ten Gaps (Priority Order)

1. **Operator notifications and reachability** (§2.2) — approvals will
   bottleneck without out-of-app notification.
2. **Dry-run/impact preview depth** (§3.2) — one bullet carrying enormous
   weight; should be its own substage.
3. **Rollback semantics uniformity** (§6.1) — browser/desktop have no rollback,
   and that's where risky actions live.
4. **Agent error remediation hints** (§3.4) — cheapest way to cut LLM cost and
   improve agent reliability.
5. **E2E encryption decision** (§5.1) — needs to be an early ADR, not a late
   hardening item.
6. **Adapter SDK + conformance suite** (§4.2) — blocks Tier-2 DCC support and
   any third-party integration.
7. **Windows/Linux host scope decision** (§1.1) — either commit to it or
   explicitly scope it out, don't leave ambiguous.
8. **Live cost visibility in Control Room** (§2.5) — operators can't make good
   approve/deny decisions without it.
9. **Multi-session supervision UI** (§2.6) — the phase talks about multi-host
   fleets (`180.C.30`) but not multi-session operator flow.
10. **Structural split of oversized stages** (§9.1) — reduces risk of
    batch-marking shortcuts.
