feat(intake): dashboard Bundles section + start_bundle skill

agents/conductors/intake/AGENTS.md

@@ -91,16 +91,16 @@ schema — light structure over free-form prose.
 | **classify** | `intake "<text>"` / `intake classify --file P` | classify one raw input; `--apply` writes the prompt |
 | **ideas** | `intake ideas` | scan `ideas.md`, propose one prompt per bullet; `--apply` writes them |
 | **census** | `intake census` | inventory every filed prompt (work-type/target/difficulty/status + hygiene flags); always read-only |
-| **dashboard** | `intake dashboard` | render the census as the Mind **task** page — picks, in flight, parked, planned, backlog, recent, epics; `--apply` writes `PyAutoMind/dashboard.md`, `--check` exits 1 on drift |
+| **dashboard** | `intake dashboard` | render the census as the Mind **task** page — picks, in flight, parked, planned, backlog, bundles, recent, epics; `--apply` writes `PyAutoMind/dashboard.md`, `--check` exits 1 on drift |
 | **formalise** | `intake formalise [prefix]` | retroactively header the prompts census flags — derive the missing fields, insert in place, prose untouched; `--apply` writes |
 | **reconcile** | `intake reconcile [prefix]` | rank backlog prompts that look already-shipped (vs the `complete/` records / `active/`), and pair live prompts that look like the same work filed twice; always read-only — retiring stays human |
 | **reconcile --repo** | `intake reconcile --repo <target> [prefix]` | **also** read the target repo's source: identifiers the prompts name that exist upstream, and lines they quote that are **gone** — the two signals that see a prompt with no Mind-side trace. Opt-in; the default path is offline |
 
 **Recent** is the one section laid out by *date* rather than by state: the 50
 newest events on the **work in hand** — issued, parked, filed — merged across
 every live bucket (the `draft/` backlog included, which is most of them:
-150 prompts against a handful of registry rows) and sitting between the Backlog
-and the Epics. Epic members stay out, as they do in every pick list on the
+150 prompts against a handful of registry rows) and sitting between the
+Bundles and the Epics. Epic members stay out, as they do in every pick list on the
 page — they are worked in order through their epic, and a Recent row hands out
 a standalone `/start_dev`. It *holds* 50
 and *shows* 10 (`RECENT_MAX` / `RECENT_PAGE`): the table is a glance, not a
@@ -112,6 +112,26 @@ none of them can answer "what has been happening?". Dates come from the
 registry key that names the event (`issued:` / `parked:` / `filed:`, PyAutoMind
 REFERENCE.md "Task dates") or a prompt's own `Issued:` header.
 
+**Bundles** read the backlog a second way — as *sessions* rather than as
+tasks — and sit directly under it. A bundle is a set of **independent** prompts
+worth running in one orchestrated session (`start_bundle`: one issue and one PR
+per member, one shared worktree per repo), which makes it the exact opposite of
+an epic: no order, no phase gate, and members stay in every pick list above,
+because a bundle is an additional VIEW and never a replacement. Two sources:
+**pinned** entries in `PyAutoMind/bundles.md` (plus any prompt whose header says
+`Bundle: <slug>`), and **auto** bundles this renderer computes — same target
+repo, no epic member, no declared `Blocked-by:`, no `human-required`, no
+`too-large`, packed under a size cap (`BUNDLE_*`: small 1 / medium 2 / large 4
+points, cap 8, at most one large, 2-4 members) in priority-then-path order.
+The proposals are then ranked (most urgent member, then biggest session, then
+slug) and only the first `BUNDLE_LIST_MAX` reach the page, with a footer saying
+what was left off and that pinning keeps it — the same pick-list-not-inventory
+rule as `PICK_LIST_MAX`. Pinned bundles are never capped.
+Auto bundles are **render-only**: recomputed on every render and never written
+back, so the nightly refresh commits no churn and a proposal that stops making
+sense stops being proposed. `Blocked-by:` reads as unresolved whatever GitHub
+says — this renderer makes no network call.
+
 Shipped work is deliberately **not** in the feed and `complete/` is never
 opened to render it: the ledger is a thousand records deep and takes ~200 a
 month, so including it made the table a list of receipts — twenty things nobody