feat(dashboard): Interventions page improvements and action-confirm modal

- Move Interventions under Workstreams in the navigator
- Add action-confirm.js: shared modal component for actions requiring a
  mandatory comment (survives live-poll re-renders, unlike inline DOM mutation)
- Wire action-confirm into Interventions (Mark done) and Decisions (Resolve)
- Fix Interventions completed section: fetch all tasks and filter client-side
  so resolved interventions (needs_human=false) still appear under Completed
- Add docs/interventions.md help page with ? button on the h1
- Replace all hardcoded "Bernd" with "human" across dashboard src and docs

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

dashboard/src/docs/interventions.md

@@ -0,0 +1,84 @@
+---
+title: Interventions — Reference
+---
+
+# Interventions — Reference
+
+The Interventions page lists every task that has been flagged `needs_human=true` — actions that only a human can take. It polls the API every 15 seconds and organises tasks into **Open** and **Completed / Cancelled** sections.
+
+---
+
+## What is a human intervention?
+
+A human intervention is a task that an agent has determined cannot proceed without direct human action — for example, approving a financial decision, confirming a legal commitment, or resolving a sensitive ambiguity. Flagging a task does not change its work status; the task continues to be `todo`, `in_progress`, or `blocked` while awaiting attention.
+
+---
+
+## KPI sidebar card
+
+| Metric | Meaning |
+|---|---|
+| **open** | Number of tasks currently flagged and not yet resolved |
+| **critical / high** | Subset with `critical` or `high` priority — shown in red when non-zero |
+| **by domain** | Top-3 domains by open intervention count |
+
+---
+
+## Open section
+
+Tasks are sorted by priority (critical → high → medium → low), then by status (blocked → in_progress → todo). Each card shows:
+
+| Element | Meaning |
+|---|---|
+| Priority badge | `critical` / `high` / `medium` / `low` |
+| Status chip | Current task status |
+| Domain | Source domain slug |
+| Workstream | Parent workstream title |
+| Action note | The `intervention_note` — what the human needs to do |
+| Task detail | Expandable `<details>` with the task title and description (shown when different from the action note) |
+
+---
+
+## Marking an intervention as done
+
+Click **Mark done** on any open card. An inline form appears requiring a **resolution comment** — a short note describing what was done. The comment is mandatory; clicking **Confirm** without entering text highlights the field in red and does nothing.
+
+Once confirmed, the API call sets `status = done`, `needs_human = false`, and replaces the action note with your resolution comment. The card moves to the **Completed / Cancelled** section on the next poll.
+
+Click **Cancel** to dismiss the form without making changes.
+
+---
+
+## Flagging and clearing interventions via MCP
+
+```
+flag_for_human(
+  task_id = "<uuid>",
+  note    = "Approve the Q2 budget allocation before the 15th"
+)
+```
+
+```
+clear_human_flag(task_id = "<uuid>")
+```
+
+`clear_human_flag` sets `needs_human = false` and preserves the `intervention_note` as a historical record. It does **not** change the task's work status. Use the dashboard's **Mark done** button when you also want to advance the status to `done`.
+
+---
+
+## Filtering by workstream
+
+`list_human_interventions(workstream_id="<uuid>")` via MCP returns only interventions for a specific workstream — useful for scoped reviews in agent sessions.
+
+---
+
+## REST endpoints
+
+| Method | Path | Effect |
+|---|---|---|
+| `GET` | `/tasks/?needs_human=true` | List all flagged tasks |
+| `PATCH` | `/tasks/{id}/` | Update status, needs_human, intervention_note |
+
+---
+
+*Interventions are never created from the dashboard — they are raised by agents during task execution and resolved by the human.*