docs(api): per-symbol pages, fuzzy search, deep token linking, hover cards

Rebuild the API Reference around one page per symbol and richer, verified content.

Pages & navigation
- One HTML page per symbol (kw-*, type-*, phase-*, screen-*, fn-*, annot-*, op-*)
  instead of a single scrolling page; namespace overview pages (ns-screen …
  ns-color) and a searchable index (api.html) with client-side fuzzy search.
- Sticky-header scroll offset (scroll-margin) so a jumped-to entry/param/color is
  never hidden, plus a flash highlight on the scrolled-to target.

Deep linking in every snippet & example
- Namespace members split: `Screen`→namespace page, `fill_rectangle`→method page;
  `Color`→palette page, `Charcoal`→its swatch — separately.
- Named arguments (`width:`) link to that parameter's anchor on the method page.
- Hover any token for a summary card built from the real API data (symbols.json).

Content & coverage
- Full authoritative surface documented from the compiler: every keyword, type,
  the 6 phases (Start/Input/FixedUpdate/Update/LateUpdate/Render, each its own
  page), all 22 annotations, namespace methods with parameter docs, builtins,
  the world_* reflection ABI, networking, operators — 155 symbols.
- Longer, clearer explanations; "model"/"model instance" terminology, not "entity";
  descriptive identifiers in every example (Position{column,row}, Velocity{delta_x,
  delta_y}, Health{current,maximum}, Player/Enemy) — no Pos/Seg/x/dx.
- Accuracy fixes from compiler ground-truth: world_count() takes no arg,
  world_query_next(property, cursor) arg order, event fields bind by name; dropped
  `when` and `module` (not in the self-hosted parser).

Tooling
- inventory.json + check.py: coverage guard (every symbol has a page), duplicate-
  token guard, and broken-link guard — fail CI so docs can't drift.
- validate.py: compiles every ```ludic example against bin/ludicc (158 compile).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/language/builtins/fn-words.md

@@ -0,0 +1,31 @@
+---
+id: fn-words
+name: words
+category: builtins
+kind: builtin
+tokens: words
+sig: words(n) -> words
+tip: Allocate a raw buffer of n 32-bit words, indexable with [i].
+order: 50
+---
+
+Allocates a raw buffer of <code>n</code> 32-bit words and returns a <code>words</code> value you can index with <code>buffer[i]</code> to read or write each word as an <code>int</code>. Use it when you want a flat, fixed-size numeric scratch array — a lookup table, a small ring buffer, or per-slot counters — without declaring a model. It is lower-level than a growable <code>[]int</code> slice: the size is chosen up front and there is no <code>push</code>. For byte-granular storage use <code>bytes</code> instead.
+
+Parameters:
+- `n` — the number of 32-bit words to allocate
+
+```ludic
+program WordScratch {
+  handler SumScores phase Start {
+    let scores = words(3)
+    scores[0] = 10
+    scores[1] = 20
+    scores[2] = 30
+    var total: int = 0
+    for index in 0 .. 3 {
+      total = total + scores[index]
+    }
+    print(total)
+  }
+}
+```