feat(v1.12.0): capture fidelity — capture_states (feedback-coverage fix)

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "ooda-loop",
   "displayName": "OODA-loop",
-  "version": "1.11.0",
+  "version": "1.12.0",
   "description": "An autonomous operations layer for your live side project. It watches, re-orients from which PRs you merge and reject, and opens small revertible PRs — bounded by a HALT file, protected paths, and a hard cost cap. Built on Boyd's OODA loop. You stay in command.",
   "author": {
     "name": "Taeil Ma",

CHANGELOG.md

@@ -8,6 +8,29 @@ independently. Bump there signals migration work for downstream projects.
 
 ---
 
+## [v1.12.0] — 2026-06-21
+
+### Added — capture fidelity (the feedback-coverage fix)
+
+Dogfooding v1.11.0 on the f1 game (10 research-grounded grade→ground→regrade
+rounds, honest re-grade 0.31 → ~0.41 stills / ~0.58 true) surfaced the next maze
+variant: **measurable improvement is bounded by FEEDBACK FIDELITY.** The 5-G
+critic captured ONE low-speed first-person cockpit frame (the car kept stopping at
+the start gantry), so a chase camera (where the clearcoat hero car + HDRI
+reflections live), the HDRI sky, speed motion-blur/FOV, and slip-angle physics —
+all real, gated, merged — were INVISIBLE to the grader. The grade stalled and the
+critic kept steering effort toward only-what-was-in-that-one-frame.
+
+- **`capture_states` (evolve 5-G + config).** A dimension is only gradeable in the
+  state(s) where it MANIFESTS; capture each dimension by driving the artifact INTO
+  those states (chase view for paint, high-speed for sense-of-speed, pitched-up for
+  sky) and pass the critic ALL frames. A state that can't be reached is a
+  capture_failure (null), never a low score — don't grade a dimension from a frame
+  that structurally can't show it.
+
+Spec/config only (verify.py unchanged at 64). plugin 1.11.0→1.12.0.
+Builds on v1.8.0 per-dimension `capture_method` + v1.11.0 reference grounding.
+
 ## [v1.11.0] — 2026-06-21
 
 ### Added — Research-Grounded OODA (the anti-maze methodology)

config.example.json

@@ -281,6 +281,8 @@
       "name": "visual_fidelity",
       "weight": 0.25,
       "capture_method": "screenshot",
+      "__capture_states_doc__": "v1.12.0 (feedback fidelity) — a dimension is only gradeable in the state(s) where it MANIFESTS; a single default frame under-credits + misdirects the loop. List the states to drive the artifact into before capturing; the critic gets ALL frames. The f1 probe earned this: a chase camera / HDRI sky / speed blur / slip-angle physics were all added but every capture was one low-speed cockpit frame, so the work was invisible and the grade stalled. A state that can't be reached = capture_failure (null), not a low score.",
+      "capture_states": ["chase view, car stationary on a straight", "chase view at ~80% top speed", "camera pitched up to show the sky/horizon"],
       "description": "3D visual quality vs SHIPPED games. Score against the reference anchors, not the artifact's past.",
       "reference": {
         "score_0.10": "flat-shaded primitive meshes, solid-colour sky, no shadows/post-processing (a 1990s look)",

skills/evolve/SKILL.md

@@ -1908,7 +1908,25 @@ for dim in rubric.dimensions:
     if not independent: dim_artifact[dim] = null (capture_failure); continue
     dim_artifact[dim] = run_protected_harness(dim.gameplay_metrics_command)  -- metrics JSON
   else if method == "screenshot":
-    dim_artifact[dim] = the shared screenshot (captured once)
+    -- CAPTURE COVERAGE (v1.12.0 — feedback fidelity). A SINGLE default frame
+    -- under-credits and MISDIRECTS: a dimension is only gradeable in the state(s)
+    -- where it MANIFESTS. The f1 probe proved this — 10 grounded rounds added a
+    -- chase camera (where the clearcoat hero car + HDRI reflections live), real
+    -- HDRI sky, speed motion-blur/FOV, and slip-angle physics, but every capture
+    -- was one LOW-SPEED COCKPIT frame facing a wall, so NONE of that work was
+    -- visible to the critic → the still-grade stalled (~0.41) while the true
+    -- quality (~0.58) was far higher, and the critic kept steering effort toward
+    -- only-what-was-in-that-frame. Rule: capture each dimension in the state(s)
+    -- where it shows. dim.capture_states (list) enumerates them, e.g.
+    --   visual_paint:  ["chase_view at speed 0", "chase_view at speed 0.8"]
+    --   sense_of_speed:["cockpit at speed 0.9 on a straight"]
+    --   sky/lighting:  ["camera pitched up to the horizon"]
+    -- Drive the artifact INTO each state (set the view/speed/scenario) before the
+    -- frame; pass the critic ALL of a dimension's frames. A dimension whose state
+    -- can't be reached is a CAPTURE_FAILURE (null), NOT a low score — never grade
+    -- a dimension from a frame that structurally cannot show it.
+    states = dim.capture_states or rubric.capture_states or ["default"]
+    dim_artifact[dim] = [ capture_in(state) for state in states ]  -- 1+ frames
   else:
     dim_artifact[dim] = run dim.capture_command (or rubric.capture_command)
 -- If a dimension's harness is MISSING entirely, score it null (capture_failure) +