docs(wiki): record measured Gemma-4-12B freeform timing on ZeroGPU

deployment-strategy.md: add "Freeform 12B timing" — reload ~3s, ~17-22 tok/s
with grammar (25-41s/call), grammar ~2x cost (16.7->35.9 on/off, intrinsic to
grammar-constrained decode), the duration=30->60 abort fix, the 1500-token >60s
edge, and that ZeroGPU aborts a 3rd GPU acquisition stacked in one demo.load
handler. Bump wrapper list to three @spaces.GPU wrappers.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

wiki/deployment-strategy.md

@@ -28,10 +28,16 @@ interprets voice edits (`interpret_genome_command`); a deterministic fallback in
 ## Runtime shape (ZeroGPU)
 
 All model calls funnel through `_blocking_completion` in
-`src/buddy_fusion/runtime.py`, exposed as two `@spaces.GPU` wrappers:
-
-- `_run_completion` at `duration=30` — warm calls are ~1–2s; a small budget
-  queues better.
+`src/buddy_fusion/runtime.py`, exposed as three `@spaces.GPU` wrappers:
+
+- `_run_completion` at `duration=30` — recipe-level Qwen-3B genome/edit calls;
+  warm calls are ~1–2s, so a small budget queues better.
+- `_run_freeform_completion` at `duration=60` (`BUDDY_FREEFORM_GPU_DURATION`) —
+  the Gemma-4-12B box-geometry calls (generate / fuse / edit). These generate
+  300–800+ tokens at ~20 tok/s and hit 25–41s total; on the old shared
+  `duration=30` they overran the budget and ZeroGPU killed them mid-decode
+  (`ggml_cuda_mul_mat_q` CUDA abort → the call dies and the user gets the
+  fallback after a full wait). See "Freeform 12B timing" below.
 - `_warmup_completion` at `duration=90` — covers the one-time cold CUDA kernel
   JIT compile.
 
@@ -86,6 +92,44 @@ is compiled to a GBNF grammar and enforced, so fuse / tweak output actually
 matches `GENOME_SCHEMA` / `GENOME_EDIT_SCHEMA`. This bug masked the real model
 behaviour; fix it before judging any model's capability.
 
+### Freeform 12B timing — measured on the Space (2026-06-15)
+
+Gemma-4-12B-it Q4_K_M on ZeroGPU, from the `[zerogpu-bench]` log lines (the
+`_blocking_completion` print of `reload` / `gen` / `tok/s`):
+
+- **reload** (rebuild `Llama` + upload ~7GB GGUF to VRAM each call): ~3s — the
+  per-call cost is small and constant; it is **not** the bottleneck.
+- **decode**: **~17–22 tok/s with the grammar on**, generating 300–800 tokens →
+  **25–41s total** per generate / fuse / edit. The wait is essentially all
+  `gen`; cutting it means cutting tokens or raising tok/s.
+- **warmup** correctly pays the one-time JIT: an 8-token warmup call logs
+  `gen≈73s` once per container, so the JIT is **not** a per-call cost afterward.
+
+**Grammar costs ~2×.** Same 400-token prompt, identical config, grammar on vs
+off (`BUDDY_GRAMMAR_BENCH=1` runs the unconstrained `[grammar-bench-off]` call):
+**16.7 tok/s on → 35.9 tok/s off**. The GBNF per-token allowed-set evaluation is
+serial CPU work that does not ride the GPU, so it roughly halves throughput.
+This is **intrinsic to grammar-constrained decode, not schema bloat** — `BOX_SCHEMA`
+is already lean (plain-int/str fields, no regex, no enum on the box path), so
+simplifying the schema buys little. The lever is to drop the grammar on the
+freeform box path and lean on `parse_json_object` + `repair_boxes` (keep it on
+the recipe `GENOME_*` path, which is fragile without it — see above). Even
+unconstrained, 35.9 tok/s is still ~3× under what a 12B Q4 should hit on the
+ZeroGPU Blackwell, so a second, harder lever (Gemma iSWA `full-size SWA cache`
+path / `flash_attn` / quant) remains.
+
+**Residual crash edge:** `duration=60` covers the observed 25–41s calls, but the
+`BUDDY_FREEFORM_MAX_TOKENS=1500` cap at ~20 tok/s is ~75s > 60s — a max-length
+generation can still overrun and abort. Lower the token cap (~900–1000, which
+also trims the typical wait) or raise the duration.
+
+**ZeroGPU aborts a 3rd GPU acquisition stacked in one request.** Inside a single
+`demo.load` handler, `warmup` (90s slot) → freeform (60s) → freeform (60s) ran
+the first two fine but ZeroGPU killed the third with `GPU task aborted`. Separate
+user-action calls (each its own request) are unaffected — so production is fine,
+but a diagnostic that fires several GPU calls from one handler must keep to ≤2
+acquisitions (why `_grammar_bench` measures only the off run).
+
 ## ZeroGPU vs CPU Basic — reference baseline
 
 Measured on the sibling project (`build-a-buddy-hf`) with the **identical**

wiki/log.md

@@ -2,6 +2,21 @@
 
 ## 2026-06-15
 
+### deployment-strategy: Freeform 12B timing measured on the Space
+
+- Added a "Freeform 12B timing" subsection: Gemma-4-12B Q4 on ZeroGPU reloads
+  ~3s and decodes ~17–22 tok/s with the grammar (25–41s per generate/fuse/edit);
+  the wait is essentially all `gen`. warmup correctly pays the one-time ~73s JIT.
+- Grammar costs ~2× (16.7 → 35.9 tok/s on/off, identical 400-token bench) and the
+  cost is intrinsic to grammar-constrained decode — `BOX_SCHEMA` is already lean
+  (no regex/enum), so the lever is dropping the grammar on the freeform box path
+  (keep it on the recipe `GENOME_*` path) + leaning on `repair_boxes`. Even
+  unconstrained, 35.9 tok/s is ~3× under the Blackwell ceiling (iSWA/flash_attn).
+- Documented the `duration=30`→`60` (`_run_freeform_completion`) fix for the
+  `ggml_cuda_mul_mat_q` mid-decode abort, the residual 1500-token > 60s edge, and
+  that ZeroGPU aborts a 3rd GPU acquisition stacked in one `demo.load` handler
+  (`GPU task aborted`). Updated the "Runtime shape" wrapper list to three.
+
 ### voxel-rendering + deployment-strategy: Figma canvas (pan + pooled live 3D) and one-call bench tweak
 
 - `voxel-rendering.md`: added "Lab canvas: pan + viewport-pooled live 3D" — tree