Bigger decks: pick up to 30 activities, built in batches

The create wizard stopped at 15 activities because the API, the prompt and
the salvage path all capped at 18: one reply must be one valid JSON object,
and past ~18 activities models truncate or drop cards. Raising the cap alone
would have shipped decks that came back short.

- The count picker offers 8 / 10 / 12 / 15 / 20 / 30; a deck's "add more" row
  gained +15. The help line says bigger decks take longer.
- /api/generate accepts up to 30 (FREEZMO_DEFAULT_ACTIVITY_COUNT now clamps
  to 30 as well) and sizes its overall deadline from the number of model
  calls instead of a fixed two.
- generateActivities splits anything above 18 evenly across calls (30 → 15 +
  15, 25 → 13 + 12) and merges them into one deck. The first call keeps the
  "add to this deck" contract; later calls are told the title, the types
  already used and the topics already covered. A normalised content signature
  drops a card that repeats an earlier one, so the same idea never ships
  twice. requested/generated still travel in the stream meta.
- phase1-test: a stub Ollama answers two 15-card batches (the second repeats
  one card on purpose) → two calls, the even split, the "already covered"
  prompt, 29 merged cards, the duplicate shipped once, title from the first
  reply, and count 31 answered 400 (59 checks).
- wizard-test: the picker offers 30, the wizard sends count: 30, and the
  30-card deck is stored and opened (99 checks).

Version 0.8.4, SW v106.

README.md

@@ -3,7 +3,7 @@
 **Turn photos of your study notes into a swipeable feed of mini study games.**
 
 Freezmo is a TikTok/Gizmo-style study app. You paste or upload photos of your notes, an AI reads
-them and builds a deck of 8–15 interactive activities — flashcards, quizzes, matching games and
+them and builds a deck of 8–30 interactive activities — flashcards, quizzes, matching games and
 more — that you swipe through vertically like a feed. Every answer gives instant feedback, XP and
 streaks, and the deck is saved in your browser so you can retry it later.
 
@@ -105,7 +105,8 @@ no database, no data leaving your machine except the requests you send to Ollama
 - Up to 8 photos + 6 documents per deck (max 12 images reach the model); photos are resized to max
   1600 px and re-encoded as JPEG before upload
 - Optional extra text (pasted notes) and a topic hint
-- Easy / Medium / Hard difficulty and 8 / 10 / 12 / 15 activities per deck
+- Easy / Medium / Hard difficulty and 8 / 10 / 12 / 15 / 20 / 30 activities per deck (a big deck
+  is built in several model calls — see the generation pipeline)
 - Fun staged loading state (`Reading your notes… → Finding the key ideas… → Building your activities…`)
 - Friendly error cards with a **Retry** button that point at Settings when a key is needed
 
@@ -782,7 +783,7 @@ Limits and behaviour worth knowing:
               ▼                                                     ▼
    ┌──────────────────────────┐                        ┌───────────────────────────┐
    │ browser: study feed      │ ◄───────────────────── │ TEXT_MODEL                │
-   │ JSON validated by zod    │   deck JSON            │ design 8–15 activities    │
+   │ JSON validated by zod    │   deck JSON            │ design 8–30 activities    │
    └──────────────────────────┘                        └───────────────────────────┘
 ```
 
@@ -794,8 +795,14 @@ Limits and behaviour worth knowing:
   with the validation error fed back to the model (`lib/ai.ts`).
 - Individual broken activities are dropped rather than failing the whole deck; if fewer than half
   the requested activities survive, the step is retried too.
+- **Bigger decks are built in batches.** One model reply has to be one valid JSON object, and past
+  ~18 activities replies get truncated or lose cards, so a count above that is split across several
+  calls (30 → 15 + 15, 25 → 13 + 12) with the same notes. Each call after the first is told the
+  deck's title, the types already used and the topics already covered, and the merged deck drops
+  cards whose content repeats (so the same idea never ships twice). The route's overall deadline
+  scales with the number of calls.
 - Requests use `stream: false` with per-request timeouts and a hard overall deadline
-  (two sequential model calls). Errors surface as friendly messages
+  (the vision step plus one model call per batch). Errors surface as friendly messages
   ("Ollama isn't reachable, check your settings") with a retry button.
 - `POST /api/generate` answers with newline-delimited JSON (`stage` / `deck` / `error` events) so the
   UI can show real progress between the vision and generation steps.
@@ -1513,6 +1520,22 @@ Things I chose while building, with the reasoning:
     into view as you type. No strings, so no i18n; the additive change sits on top of the merged
     scene-heat rework unchanged.
 
+112. **Bigger decks are one request but several model calls.** The create wizard topped out at 15
+     activities (`COUNT_OPTIONS`) because the API, the prompt and the salvage path all capped at 18:
+     one reply has to be one valid JSON object, and past that size models start truncating or
+     quietly dropping cards. Raising the cap alone would have shipped decks that came back short.
+     The picker now offers 8 / 10 / 12 / 15 / 20 / 30 (and a deck's "add more" row gained +15); the
+     route accepts up to 30; and `generateActivities` splits anything above 18 evenly across calls
+     (30 → 15 + 15), merging them into one deck. The first call keeps the existing "add to this
+     deck" contract; later calls are told the title, the types already used and the topics already
+     covered, and a normalised content signature (`activitySignature`) drops a card that repeats an
+     earlier one — the stub in phase 1 returns the same first card in both batches to pin exactly
+     that (29 from two 15-card replies). The route's overall deadline now scales with the number of
+     calls, and the requested/generated pair still travels in the stream's meta. Verified by the
+     phase-1 suite (59 checks: the two calls, the even split, the "already covered" prompt, the
+     merge + dedupe, and a 31-card request answered 400) and the wizard suite (99: the picker
+     offers 30, the wizard sends `count: 30`, and the 30-card deck is stored and opened).
+
 111. **The model dropdown was clipped by the card it lived in.** "The model selection in teach it
      back is kinda broken at least on desktop": the pickers' desktop panels were
      `position: absolute` children of whatever container held their chip. In chat that is the