feat(api): add structured schedule plan with topics, problems, gate, sweep and reserved problem handling

apps/api/README.md

@@ -14,8 +14,12 @@ Live at `https://srs-api.prdlk.workers.dev`.
   labels, milestones. `catalog.ts` reconciles them into D1 nightly (and via
   `/admin/reconcile`); it recomputes from scratch, never invents rows, and
   never overwrites SRS-owned columns (`stage`, `next_review`).
-- **The repo owns the schedule** — `data/schedule.json`, human-edited,
-  bundled at deploy. Git history is its audit log.
+- **The repo owns the plan** — `data/schedule.json`, human-edited,
+  bundled at deploy. Git history is its audit log. Each date is a
+  `ScheduledDay`: `topics` (one or more topic issues), `problems` (a set day,
+  fixed numbers), `gate` (Saturday's blind set, verbatim), or `sweep` (redo
+  every miss, nothing new). Numbers a `gate` or `problems` entry names are
+  `RESERVED`: the drill and gate samplers never draw them.
 - The GitHub Project mirror writes `Target Date` and nothing else: it is
   best-effort, so failures log warnings and never block a D1 write, and topic
   rows are never written. SRS stage and first-attempt result stay in D1 only —
@@ -47,7 +51,10 @@ raw `Date` math.
 Blind drills draw only from topics **already learned**: scheduled in an
 earlier week (the current week's optional pool is reserved for Saturday's
 gate) and with at least one core problem on the ladder — a skipped learning
-day never feeds drills just because its calendar week lapsed.
+day never feeds drills just because its calendar week lapsed. A Saturday
+whose plan fixes the `gate` set uses that list as-is; every number must be a
+catalog row (a `set:optional` sub-issue of its topic), or the review issue is
+not created — a `/done` on an uncatalogued number could never be logged.
 
 ## Routes