fix(m6): post-review pass — cache prefix, snapshot lock, perm-before-parse, LIKE escape

Addresses spec-reviewer + code-reviewer feedback on the M6 bundle:

Critical:
- frontend/src/lib/missions.ts: add `listPrefix()` so TanStack invalidation
  catches every filtered list variant; the previous `list()` returned
  `['missions','list',{}]` and only matched the exact empty-filter cache,
  leaving filtered tables stale after create/transition/delete.
- backend/app/services/missions.py: acquire the same per-scenario
  `pg_advisory_xact_lock` key used by `set_scenario_tests` before
  snapshotting; without it a concurrent M5 reorder could freeze a torn
  snapshot under READ COMMITTED. Sorted by key to avoid deadlocks with
  another snapshotter.

Important:
- backend/app/api/missions.py: `@require_perm("mission.update",
  "mission.archive")` on the transition endpoint so users without either
  perm get 403 before the body is parsed (no shape leak via 400).
- backend/app/services/missions.py: escape `%` / `_` / `\` in user-typed
  `q` / `client` LIKE search; users can no longer trigger wildcard
  semantics by typing literal `%`. Added `escape='\\'` arg on every .like().
- backend/app/services/missions.py: filter `MissionTest.deleted_at` and
  `MissionScenario.deleted_at` in the list-item and detail counts so M7+
  soft-deletes don't drift the totals silently.

Nits:
- backend/app/api/users.py: order `/users/roster` by email for stable
  rendering + deterministic e2e selectors.
- frontend/src/pages/MissionDetailPage.tsx: distinct accent per
  transition target (cyan/orange/green/teal) matching the status legend.
- e2e/tests/m6-missions.spec.ts: switch fragile `getByRole(name=/In
  Progress/i)` to the stable `mission-transition-in_progress` data-testid.

New tests:
- test_create_mission_rejects_soft_deleted_scenario
- test_transition_perm_gate_runs_before_payload_parse
- test_search_treats_wildcards_as_literals

Suite: 106 pytest passing (was 103), 43 Playwright passing.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -4,6 +4,17 @@ All notable changes to this project will be documented here. Format: [Keep a Cha
 
 ## [Unreleased]
 
+### Fixed (post-M6 review pass — spec-reviewer + code-reviewer)
+- **SPA cache invalidation only refreshed the empty-filter list** (`frontend/src/lib/missions.ts:136`): `missionKeys.list()` returns `['missions','list',{}]`. TanStack v5's `invalidateQueries({queryKey})` is prefix-based, but `{}` is treated as an atomic final element — so create / transition / delete called with that key only invalidated the *exact* empty-filter list, leaving any filtered variant stale until manual refetch. Added `missionKeys.listPrefix()` returning `['missions','list']` and switched all three mutation `onSuccess` paths to it.
+- **Snapshot lacked the per-scenario advisory lock** (`backend/app/services/missions.py:467`): a concurrent `PUT /scenario-templates/{id}/tests` (M5 reorder, which deletes-then-reinserts join rows) running while `_snapshot_scenarios` walked `sc.tests` could freeze a torn snapshot — `selectinload` re-queries under READ COMMITTED so a partial view was possible. Added `_lock_scenario_ids_for_snapshot` that acquires the same `pg_advisory_xact_lock` key used by `set_scenario_tests` (blake2b digest of the scenario UUID, sorted to avoid deadlocks). Snapshot and reorder now serialise per scenario.
+- **Transition endpoint leaked its body shape via 400 before the perm gate** (`backend/app/api/missions.py:441`): a user without `mission.update` or `mission.archive` POSTing `{"status":"x"}` got a Pydantic 400 instead of 403. Added `@require_perm("mission.update", "mission.archive")` so the gate fires before the parse; the inner refinement still enforces the per-target perm. Test `test_transition_perm_gate_runs_before_payload_parse`.
+- **LIKE wildcards in user-typed search were honoured as SQL wildcards** (`backend/app/services/missions.py:632,637`): `?q=%` matched every mission. Added `_escape_like` that pre-escapes `%`, `_`, `\` and a matching `escape='\\'` argument on every `.like(...)` call. Test `test_search_treats_wildcards_as_literals`.
+- **Counts ignored soft-deleted mission children** (`backend/app/services/missions.py:587,597`): `tests_count` and the detail view summed `len(sc.tests)` without filtering `MissionTest.deleted_at`. Harmless today (M6 doesn't soft-delete mission tests), but would drift silently once M7+ surfaces `state=skipped/blocked`. Added the filter in both `_to_list_item` and `_scenario_views`.
+- **`/users/roster` was unordered** (`backend/app/api/users.py:73`): the wizard's member list shuffled rows on every refetch. Sorted by `email` for predictable rendering + stable e2e selectors.
+- **Frontend transition button accent collapsed `in_progress` and `completed` into one colour** (`frontend/src/pages/MissionDetailPage.tsx:97`): both rendered cyan, so the status legend in the list didn't match the transition button. Added a `TRANSITION_BUTTON_ACCENT` map mirroring `MISSION_STATUS_ACCENT` (cyan/orange/green/teal).
+- **Soft-deleted source scenario was a silent foot-gun**: `_load_scenario_templates_for_snapshot` already rejected it, but no test pinned the behaviour. Added `test_create_mission_rejects_soft_deleted_scenario` so future refactors can't regress to "freeze a tombstoned scenario into a fresh mission".
+- **E2E wizard assertion used `getByRole('button', { name: /In Progress/i })`** (`e2e/tests/m6-missions.spec.ts:287`): the accessible name is `→ In Progress` and the arrow Unicode is brittle. Switched to `getByTestId('mission-transition-in_progress')`.
+
 ### Added — M6 (Missions & snapshot)
 - **CRUD `missions`** (`app/services/missions.py` + `app/api/missions.py`):
   - Fields: name, client_target, date_start, date_end, status (`draft/in_progress/completed/archived`), description (markdown), visibility_mode (frozen to `whitebox` in v1).