Add a repeatable screenshot capture and compare smoke

Both capture helpers now accept --out=res://docs/screenshots, and
tools/check_screenshots.sh renders the staged set (using xvfb-run when headless)
before running screen_designs.py compare. Record the judge result: 7 match,
4 review, 0 diff. Headless 226/226 and the smoke both pass in a clean worktree.

MEMORY.md

@@ -49,13 +49,26 @@ Session memory for durable decisions, state, and gotchas. Keep concise, remove o
 
 ## Handoff (read this first)
 
-- **Visual-alignment pass planned (2026-11-01):** `TODO.md` gains a "Visual
-  alignment with rendered mockups" section. Goal: restructure generated-image
-  storage, move the reference mockups out of `docs/screens/designs/` into a
-  runtime-accessible root (e.g. `assets/images/`), load the proposed UI/scene
-  art at runtime (start menu, selection screens, HUD, panels), and make the game
-  match the mockups — verified by `screen_designs.py compare` + `judge`. Not
-  started; no code moved yet.
+- **Status + active passes (2026-11-01, moved from `TODO.md`):** v0.6.0 shipped — the Blender
+  character/gear/prop library is integrated, AO maps are baked and wired, and the environment
+  textures are lifted to 1024² with normal maps (see `DONE.md`). Headless **222/222**, Python
+  **81**; the scheduled presentation pass is complete. Two passes remain in `TODO.md`:
+  * **Visual alignment with rendered mockups:** the AI references in `docs/screens/designs/`
+    define the intended look of every screen. The pass reorganizes generated-image storage,
+    loads the proposed UI/scene art at runtime (`src/game/ui/art.gd`, `assets/images/`), and
+    aligns the selection screens, HUD, and scene visuals with those mockups — verified with
+    `screen_designs.py compare` + `judge`.
+  * **Realism pass** (plan: `PLAN.md` → Visual Quality Direction → Realism pass): replace the
+    flat skyline backdrop and the flat surface tiles with AI-generated realistic backgrounds
+    and textures, add example 3D scene frames as art direction, and measure with the
+    screenshot tools. Rule: never overwrite an existing image — new outputs get new paths
+    (`next_free_path` -vN suffix). Raw → `assets/textures/ai/v2/raw/` (gitignored); candidates
+    → `assets/textures/ai/game_ready/v2/` (gitignored pool, only validated picks tracked,
+    ADR-0014); scene refs → `docs/screens/concepts/`; trials → `docs/screens/realism/<run>/`.
+    Realism report 2026-11-01-v2 mean judge scores: backdrop 2.6 → 3.6 (v2 day + night
+    skyline) → **4.0** (v2 day + dusk skyline for the Evening phase); texture detail 1.2 →
+    1.6–1.75; overall 1.6 → 2.0. Scores swing ±1–2 between runs (trend only); the
+    `characters` score moved 3.0 → 1.0 with identical NPCs, showing the noise.
 
 - **Phase 2 — AI texturing (this agent, DONE 2026-11-01).** Phase 1 assets were
   already in the tree (untracked) and are preserved. Generated six seamless AI
@@ -138,22 +151,36 @@ Session memory for durable decisions, state, and gotchas. Keep concise, remove o
   flat plane, so avoidance (not a baked mesh) routes guests around props. Placements live in
   `PROP_PLACEMENTS` with `placement_problems()` validation; variants come from `variant_for(index)`.
   `smoke_props.gd` must be windowed because the guard reads captured-mouse input.
-- **Realism pass tooling (2026-11-01):** `tools/scene_concepts.py` (`list`/`generate`/`process`/
-  `audit`/`realism`) never overwrites an image (`next_free_path` adds `-vN`; `process` skips
-  existing targets). Raw AI output: `assets/textures/ai/v2/raw/` (gitignored). Candidate pool:
-  `assets/textures/ai/game_ready/v2/` (gitignored except the validated picks un-ignored in
-  `.gitignore`: skyline day-v4/dusk-v3/night-v4 plus brick/pavers/asphalt/asphalt_wet + normals).
-  Concepts: `docs/screens/concepts/` (committed, `.gdignore`, targets in its README). Runs:
-  `docs/screens/realism/<run>/` (PNG captures gitignored; `report*.{json,md}` + `sheet.jpg`
-  committed). `src/test/capture_realism.gd` swaps textures at runtime by the material's
-  *original* albedo path (`--swap/--night-swap/--tint/--normal=OLD=NEW`) and writes to
-  `--out=res://...`. Changing game code is not needed for an A/B test. Gotchas:
-  gpt-5-image-mini honours `background=transparent` for the 3:2 skyline requests, but it writes
-  garbled "HOLLYWOOD" lettering unless the prompt forbids signs. Skyline strips are 2:1 to match
-  one UV repeat of the ring. `elapsed_s = 620` gives the *Evening* phase (pink sky), not night.
-  The gemini-flash-lite realism scores are noisy (±1–2), and the judge sometimes returns a 400
-  "API key not valid" from the provider, which the one retry covers. None of the v2 assets are
-  wired into the game yet.
+- **Realism pass (2026-11-01, completed + adopted):** `tools/scene_concepts.py`
+  (`list`/`generate`/`process`/`promote`/`audit`/`realism`) never overwrites an image
+  (`next_free_path` adds `-vN`; `process` skips existing unless `--new`; `promote` keeps an
+  existing pick). Raw AI output: `assets/textures/ai/v2/raw/` (gitignored). Candidate pool:
+  `assets/textures/ai/game_ready/v2/` (gitignored) and the **tracked runtime assets** are
+  `.../v2/picks/{skyline_day,skyline_dusk,skyline_night,brick_facade,sidewalk_pavers,asphalt,
+  asphalt_wet,awning_canvas}.png` (+ `_normal`). Concepts `docs/screens/concepts/` (committed,
+  `.gdignore`, targets in its README). Runs `docs/screens/realism/<run>/` (PNG captures
+  gitignored; `report.{json,md}` + `sheet.jpg` committed).
+  `src/test/capture_realism.gd` runs five stages (day 0 s / evening 620 s / late 760 s /
+  joke_shop 700 s / studio 700 s; `--stages=` subset) and swaps textures by the material's
+  *original* albedo path. **Gotcha:** in the SceneTree script, do not overwrite `_phase` after
+  `_begin_stage()` sets `done`, or the run loops forever between stage/shoot (`_queue` empty).
+  **Gotcha:** `elapsed_s = 620` is *Evening* (dusk), 700+ is Late. Godot 4.3
+  `StandardMaterial3D` has no `fog_enabled`; the skyline is a `ShaderMaterial` with
+  `render_mode unshaded, cull_disabled, fog_disabled` + `ALPHA_SCISSOR_THRESHOLD` so the
+  panorama's own haze shows. Backdrop plane is 90 m tall with 2 UV repeats (~200×90 m per
+  repeat, near the 2:1 panorama aspect). Legacy `assets/textures/skyline.png` is no longer
+  loaded but stays on disk for the `test_world` asset check.
+  **Phase look:** day sun 0.62 / exposure 0.6 / ambient 0.22; evening sun 0.12 / exposure
+  0.72 / dark violet sky; late exposure 0.95 / ambient 0.10 with the wet asphalt street and a
+  bright skyline tint (1.5) so the night windows read. Stripes: `stripe_period()` (detrended
+  autocorrelation) + `align_stripes()` (resample to whole periods) replace the edge crossfade
+  for `awning_canvas`; `audit_striped_tile()` compares the wrap step with the harshest interior
+  step.
+  **Judge:** `google/gemini-2.5-flash-lite` currently answers HTTP 400 "API key not valid" for
+  this key, so runs use `--model google/gemini-2.5-flash`; scores are noisy (±1–2), so
+  `realism --repeats 3` averages and `summarize()` reports a like-for-like mean over the shots
+  every set captured. Final: overall 0.8→1.0, backdrop 0.6→1.6, metric gap 0.444→0.348.
+  `screen_designs.py judge` gained one retry for the same transient 400.
 - **Blender asset library (v0.6.0, integrated 2026-11-01):**
   `tools/blender_assets.py` (`build`/`render` in Blender; `audit`/`sheet`/`list` stdlib) writes
   `assets/models/blender/{npc,npc_gear,props}/*.glb`, `manifest.json` (integration source of
@@ -212,6 +239,7 @@ Session memory for durable decisions, state, and gotchas. Keep concise, remove o
 - Radio, wand, and body-camera timeline panels use their reference art as backdrops; forms sit on transparent, padded panels over the covered mockup controls.
 - HUD target names use a boxed chip under the crosshair; escalation has a five-segment bar below the text label.
 - Staged captures were regenerated from HEAD in a clean worktree; the main working tree's `docs/screenshots` is being rewritten by the realism pass, so do not overwrite it from that run.
+- `tools/check_screenshots.sh` is the repeatable capture+compare smoke; both capture scripts accept `--out=res://docs/screenshots`. Judge verdicts on the staged set were 7 match / 4 review / 0 diff.
 
 ## Current State