feat(copilot): autopilot UX polish + skills/followups LD kill-switches + unified Scheduled page (#13202)

## Why

Three things were left over after the autopilot consolidation shipped
(#13190 + #13195):

1. **Naming drift** — the briefing pill said "Copilot library" even
though the product is autopilot.
2. **Per-link hide on 0** — "1 skill · 0 follow-ups" reads as broken UX.
Either count being zero should hide just that link.
3. **No kill-switch** — skills + scheduled-followups inject a
`<available_skills>` / `<session_context>` block on every turn and gate
the four MCP tools. If something regresses we need to disable them
per-user without a redeploy.

Plus one piece of unification a sleep-late-Friday user asked for:
`/library/followups` should show **both** copilot followups AND
recurring agent schedules, since they're the same APScheduler primitive
with a `kind` discriminator and the UX of having two separate pages is
silly.

## What

### LD kill-switches (default-on)

Two new server-side LaunchDarkly flags, evaluated per-user, default
`true`:

- **`copilot-skills`** — gates the four MCP tools (`store_skill` /
`read_skill` / `delete_skill` / `list_skills`) and the per-turn
`<available_skills>` block. When off, `build_skills_context` returns
empty (no list query, no Redis hit) and every tool returns a structured
`feature_disabled` error.
- **`copilot-scheduled-followups`** — gates the `schedule_followup` MCP
tool and the `pending_followups` line inside `<session_context>`. When
off, the tool returns `feature_disabled` and `build_session_context`
collapses to the bare `session_id` line WITHOUT calling the scheduler
RPC.

Local override pattern unchanged: `FORCE_FLAG_COPILOT_SKILLS=false` /
`FORCE_FLAG_COPILOT_SCHEDULED_FOLLOWUPS=false` for dev.

### Briefing pill (`CopilotLibrarySummary`)

- Label: `Copilot library` → **`Autopilot library`**.
- Counts are independent now: shows `N skill(s)` only if N > 0, shows `M
scheduled` only if M > 0, drops the `·` separator when only one is
shown. Still hides the whole pill when both are 0 (no degenerate state).
- Scheduled count now sums **followups + graph schedules** so it mirrors
what `/library/followups` displays.

### Unified Scheduled page (`/library/followups`)

Route name unchanged for backward compat, but the page is now
"Scheduled":

- Title `Copilot follow-ups` → **`Scheduled`**.
- `useFollowupsPage` fetches BOTH `listCopilotFollowupSchedules` AND
`getV1ListExecutionSchedulesForAUser`, merges into a discriminated
`ScheduleItem[]`, sorts by `next_run_time` ascending.
- New `GraphScheduleListItem` component renders graph rows with an
emerald "Agent run" badge, agent name, cron label, View dialog showing
graph_id + schedule name, and a Delete flow that invalidates the
graph-schedules query.
- Existing `FollowupListItem` (copilot_turn) untouched — just rendered
as one branch of the polymorphic list.
- Empty-state copy: "No copilot follow-ups" → "Nothing scheduled" +
mention both surfaces.

## How (highlights)

- **Skills helper**: `is_skills_feature_enabled(user_id)` returns `True`
for anonymous (no user_id) so we don't surprise-break unauthenticated
paths; otherwise reads the flag with `default=True`.
- **Followups helper**: `is_followups_feature_enabled(user_id)` same
shape, lives in `session_context.py` so the gating function ships with
the consumer.
- **Cost half of the kill-switch**: when off, scheduler RPC / list-files
/ Redis fetches are all skipped — flipping the flag also drops the
per-turn cost to zero, not "still calls but ignores result".
- **Frontend merge over backend union**: kept the two existing typed
REST endpoints (`/schedules` returns `list[GraphExecutionJobInfo]`,
`/schedules/followups` returns `list[CopilotTurnJobInfo]`) and stitched
them on the client. A unified `/schedules/all` endpoint with a
discriminated union would have been cleaner but required a new model +
openapi regen for marginal benefit.

## Test coverage

- **Backend** (102 passed):
- `build_skills_context_empty_when_flag_disabled` — short-circuit
verified
- `build_skills_context_normal_when_flag_enabled` — happy path still
works
- `store_skill_returns_feature_disabled_when_flag_off` + `list_skills_…`
— tool gating
- `build_session_context_skips_scheduler_when_followups_disabled` —
verifies the scheduler RPC is NOT called when off (via
`AssertionError(...)` side-effect on the mock)
- `schedule_followup returns_feature_disabled_when_flag_off` — tool
gating + scheduler spy
- **Frontend**: existing skills + followups page suites should still
pass post-rename; will run via CI.
- **Live functional /pr-test**: planned on dev preview after the
redeploy that pulls this commit.

## Out of scope

- Unifying the MCP tool surface (one `schedule(kind=...)` instead of two
tools) — refactor deferred.
- New `/schedules/all` REST endpoint — frontend stitching is enough
today.
- Renaming the `/library/followups` route — backward compat with
existing deep-links from the briefing pill etc.

## Checklist

- [x] Code follows project style guidelines
- [x] Self-review performed
- [x] Tests added for new behaviors (flag-off paths)
- [x] No new warnings
- [x] Authorization unchanged — endpoints + helpers still scope by
`user_id` from auth
- [ ] CI green
- [ ] `/pr-test` on dev-preview validates pill + Scheduled page + flag
kill-switch