Merge pull request 'test(M00): record final paired evidence' (#13) from work/sindo-main-codex/m00-final-paired-evidence into master

Reviewed-on: #13

coordination/claims/M00-QAR-CAMERA-FRAMING-REFINEMENT-001.md

@@ -0,0 +1,78 @@
+# M00-QAR-CAMERA-FRAMING-REFINEMENT-001 — Goldshire pitch/FOV refinement
+
+<!-- OPENWC_CLAIM:M00-QAR-CAMERA-FRAMING-REFINEMENT-001:sindo-main-codex:2026-07-14 -->
+
+## Ownership
+
+- Target: M00
+- Program: QAR
+- Owner/Agent ID: sindo-main-codex
+- Branch: `work/sindo-main-codex/m00-camera-framing-refinement`
+- Lease expires UTC: 2026-07-14
+- Integrator: milestone integrator
+
+## Outcome
+
+Jointly refine Goldshire checkpoint pitch and FOV against a WMO-ready build 12340 reference.
+
+## Non-goals
+
+- Changing runtime player camera defaults.
+- Changing manifest values without an interior optimum and human approval.
+- Optimizing yaw beyond the ready-scene coarse result.
+
+## Paths
+
+- Exclusive: joint FOV option in camera-pose sweep and framing evidence
+- Shared/hotspots: camera-pose operational documentation and renderer module findings
+- Generated/ignored: candidate PNGs and ranking JSON under `user://`
+
+## Contracts and data
+
+- Public API/events: additive `-CameraFovValues`; scalar `-CameraFov` remains compatible
+- Schema/format version: ranking schema remains version 1 with existing per-candidate FOV field
+- Migration/compatibility: scalar sweeps retain existing candidate paths
+- Consumers: M00 fidelity workflow
+
+## Dependencies
+
+- Requires: WMO-ready 8-second capture contract
+- Blocks: Goldshire manifest framing decision
+- External state: private build 12340 reference remains outside Git
+
+## Verification
+
+- Commands: joint-grid plan regression, real ready-scene sweep, comparator self-test, repository gates
+- Fixtures: private Goldshire Inn reference
+- Fidelity evidence: joint pitch/FOV ranking and human inspection
+- Performance budget: offline diagnostic
+
+## Documentation deliverables
+
+- Inline public API docs: PowerShell parameter
+- Module specification: framing finding
+- Data-flow diagram: FOV joins pose grid
+- Sequence/state/dependency diagrams: unchanged
+- Source map/status updates: operational guide
+
+## Simplicity and naming
+
+- Important names introduced: `CameraFovValues`, `effectiveCameraFovValues`
+- Simplest considered solution: one additional loop around the existing grid
+- Rejected complexity/abstractions: generic optimizer
+- Unavoidable complexity and justification: projection and pitch jointly affect framing
+- Measured optimization evidence: bounded 1x5x3 grid
+
+## Status
+
+- State: ready-for-review
+- Done: additive joint FOV grid, compatibility-preserving paths, plan regression, ready-scene 15-candidate ranking and human inspection
+- Next: add approved landmark/region scoring before any manifest camera calibration
+- Blocked by:
+
+## Handoff
+
+- Commit: this work-package commit
+- Results: numeric best FOV 62/yaw 10/pitch -25 at mean 0.078843 and ratio 0.667721 was visually rejected as grass-dominated; FOV 38/yaw 10/pitch -10 better matched building scale at mean 0.084220
+- Remaining risks: full-frame metric does not preserve landmark framing; no manifest values were changed
+- Documentation updated: `docs/CAMERA_POSE_SWEEP.md`, `docs/modules/world-renderer.md`

coordination/claims/M00-QAR-FINAL-PAIRED-EVIDENCE-001.md

@@ -0,0 +1,78 @@
+# M00-QAR-FINAL-PAIRED-EVIDENCE-001 — Final paired evidence audit
+
+<!-- OPENWC_CLAIM:M00-QAR-FINAL-PAIRED-EVIDENCE-001:sindo-main-codex:2026-07-14 -->
+
+## Ownership
+
+- Target: M00
+- Program: QAR
+- Owner/Agent ID: sindo-main-codex
+- Branch: `work/sindo-main-codex/m00-final-paired-evidence`
+- Lease expires UTC: 2026-07-14
+- Integrator: milestone integrator
+
+## Outcome
+
+Produce one final WMO-ready paired report for all available build 12340 references and reconcile the remaining M00 acceptance gaps.
+
+## Non-goals
+
+- Achieving renderer parity inside M00.
+- Calibrating unknown manual camera parameters.
+- Marking the target DONE; only the milestone integrator may do so.
+
+## Paths
+
+- Exclusive: paired-run exit handling and final evidence record
+- Shared/hotspots: renderer baseline/module findings
+- Generated/ignored: final captures and reports under `user://`
+
+## Contracts and data
+
+- Public API/events: visual comparison exit 0/1 both count as completed report generation
+- Schema/format version: unchanged
+- Migration/compatibility: comparator infrastructure errors remain fatal
+- Consumers: M00 integrator
+
+## Dependencies
+
+- Requires: merged region diagnostic and five private reference images
+- Blocks: integrator decision on M00 completion
+- External state: private images remain outside Git
+
+## Verification
+
+- Commands: complete paired baseline run, report audit, repository gates
+- Fixtures: five build 12340 references
+- Fidelity evidence: cold/warm comparison metrics and human gap classification
+- Performance budget: existing M00 capture budgets
+
+## Documentation deliverables
+
+- Inline public API docs: PowerShell comparison step semantics
+- Module specification: final evidence status
+- Data-flow diagram: unchanged
+- Sequence/state/dependency diagrams: unchanged
+- Source map/status updates: baseline acceptance summary
+
+## Simplicity and naming
+
+- Important names introduced: `Invoke-GodotComparisonStep`
+- Simplest considered solution: allow only comparator exit codes 0 and 1
+- Rejected complexity/abstractions: suppressing all process failures
+- Unavoidable complexity and justification: threshold exceedance is evidence, not infrastructure failure
+- Measured optimization evidence: not applicable
+
+## Status
+
+- State: ready-for-review
+- Done: paired-run exit contract fixed; full baseline and visual capture completed; ten pairs audited; final acceptance document written
+- Next: integrator decides animated/sky reference policy and GUI RID-risk disposition
+- Blocked by:
+
+## Handoff
+
+- Commit: this work-package commit
+- Results: compared=10, failed=10, missing=0; mean error 0.066892..0.176952; cold/warm gaps stable; full workflow returned success with threshold evidence preserved
+- Remaining risks: no build 12340 animated-M2 or sky-transition reference; GUI renderer RID/resources still reported at exit
+- Documentation updated: `docs/M00_FINAL_PAIRED_EVIDENCE.md`, `docs/modules/world-renderer.md`

coordination/claims/M00-QAR-LANDMARK-REGION-DIFF-001.md

@@ -0,0 +1,78 @@
+# M00-QAR-LANDMARK-REGION-DIFF-001 — Landmark-region visual diff
+
+<!-- OPENWC_CLAIM:M00-QAR-LANDMARK-REGION-DIFF-001:sindo-main-codex:2026-07-14 -->
+
+## Ownership
+
+- Target: M00
+- Program: QAR
+- Owner/Agent ID: sindo-main-codex
+- Branch: `work/sindo-main-codex/m00-landmark-region-diff`
+- Lease expires UTC: 2026-07-14
+- Integrator: milestone integrator
+
+## Outcome
+
+Rank camera candidates inside an approved common screen region so unrelated grass/sky coverage cannot dominate landmark framing.
+
+## Non-goals
+
+- Automatic feature detection or image warping.
+- Different crops for reference and candidate.
+- Changing manifest camera values without human approval.
+
+## Paths
+
+- Exclusive: comparator common-region option and sweep pass-through
+- Shared/hotspots: camera-pose documentation and renderer fidelity findings
+- Generated/ignored: local ROI sweep images/reports
+
+## Contracts and data
+
+- Public API/events: additive comparator `--region x,y,width,height` and sweep `-ComparisonRegion`
+- Schema/format version: report/ranking schema remains version 1 with additive region metadata
+- Migration/compatibility: omitted region preserves full-frame behavior
+- Consumers: M00 paired camera calibration
+
+## Dependencies
+
+- Requires: ready-scene joint pitch/FOV sweep
+- Blocks: landmark-weighted Goldshire framing decision
+- External state: private build 12340 reference
+
+## Verification
+
+- Commands: comparator synthetic ROI regression, invalid-bound regression, sweep plan, bounded real ROI sweep, repository gates
+- Fixtures: synthetic 2x2 images and private Goldshire reference
+- Fidelity evidence: facade-region ranking plus human inspection
+- Performance budget: region comparison visits only selected pixels
+
+## Documentation deliverables
+
+- Inline public API docs: comparator usage and PowerShell parameter
+- Module specification: ROI finding
+- Data-flow diagram: approved region into comparator
+- Sequence/state/dependency diagrams: unchanged
+- Source map/status updates: camera-pose guide
+
+## Simplicity and naming
+
+- Important names introduced: `ComparisonRegion`, `comparison_region`
+- Simplest considered solution: one common integer rectangle
+- Rejected complexity/abstractions: masks, feature detectors, separate movable crops
+- Unavoidable complexity and justification: full-frame color error is biased by large terrain regions
+- Measured optimization evidence: bounded pixel loop
+
+## Status
+
+- State: ready-for-review
+- Done: common ROI parser/validation/reporting, synthetic inclusion/exclusion regression, sweep pass-through, 15-candidate Goldshire rescoring and human inspection
+- Next: define explicit approved landmark coordinates or mask; do not calibrate manifest from the common rectangle alone
+- Blocked by:
+
+## Handoff
+
+- Commit: this work-package commit
+- Results: ROI [300,100,1400,650] ranked FOV 38/yaw 10/pitch -15 first at mean 0.081777 and ratio 0.697185; visually the facade remains oversized and cropped
+- Remaining risks: rectangle includes non-semantic road/terrain pixels and cannot prove landmark alignment
+- Documentation updated: `docs/CAMERA_POSE_SWEEP.md`, `docs/modules/world-renderer.md`

docs/CAMERA_POSE_SWEEP.md

@@ -14,11 +14,12 @@ This offline M00 diagnostic recovers reproducible checkpoint framing when an ori
   -Checkpoint elwynn_adt_boundary `
   -YawOffsets -15,-10,-5,0,5,10,15 `
   -PitchOffsets -10,-5,0,5,10 `
+  -CameraFovValues 38,50,62 `
   -ViewportWidth 1280 -ViewportHeight 960 `
   -WaitSeconds 8
 ```
 
-Each candidate receives its own output directory and comparison report. The viewport must exactly match the reference image dimensions; the runner fails on `size_mismatch` instead of inventing a score. The default 8-second wait matches the M00 manifest and is a readiness precondition: lowering it requires evidence that the checkpoint's WMO/M2 content has already appeared. `ranking.json` sorts candidates by mean perceptual error and then changed-pixel ratio. Use `-PlanOnly` to validate the Cartesian grid and output paths without rendering. Run a coarse grid first, then a finer grid around the best candidate.
+Each candidate receives its own output directory and comparison report. `-CameraFovValues` adds FOV as another bounded grid dimension; omit it to retain the scalar `-CameraFov` workflow and legacy candidate paths. The viewport must exactly match the reference image dimensions; the runner fails on `size_mismatch` instead of inventing a score. The default 8-second wait matches the M00 manifest and is a readiness precondition: lowering it requires evidence that the checkpoint's WMO/M2 content has already appeared. `ranking.json` sorts candidates by mean perceptual error and then changed-pixel ratio. Use `-PlanOnly` to validate the Cartesian grid and output paths without rendering. Run a coarse grid first, then a finer grid around the best candidate.
 
 ```mermaid
 flowchart LR
@@ -47,4 +48,16 @@ These values are invalid camera-pose evidence because the sweep used a 2-second
 
 The ready-scene 3x3 rerun at FOV 62 degrees ranked `(yaw=10, pitch=-10)` first with mean error `0.087952` and changed-pixel ratio `0.665527`, compared with `0.101402`/`0.702489` at zero offsets. Human inspection confirms that the inn is present and the direction is plausible. Pitch remains on the tested boundary and the building scale differs from the reference, so these offsets are coarse evidence only; a joint pitch/FOV refinement is required before changing manifest camera values.
 
+## Goldshire joint pitch/FOV refinement — 2026-07-12
+
+A WMO-ready joint grid fixed yaw at 10 degrees and evaluated pitch `[-25, -20, -15, -10, -5]` across FOV `[38, 50, 62]`. Full-frame error ranked FOV 62/pitch -25 first at `0.078843` with changed-pixel ratio `0.667721`. Human inspection rejects it as a manifest calibration: the view is dominated by grass and crops the inn. FOV 38/pitch -10 is closer to the reference building scale but ranks only `0.084220`, and horizontal landmark alignment remains different.
+
+The grid therefore has no approved joint optimum. Full-frame color error continues to trade landmark framing against large terrain regions. The manifest remains unchanged. The next registration method must score an approved inn/road/tree landmark region or explicit landmark coordinates rather than the entire image.
+
+## Common landmark region — 2026-07-12
+
+The comparator accepts an optional common screen rectangle through `--region x,y,width,height`; the sweep runner exposes it as `-ComparisonRegion x,y,width,height`. Coordinates are integer pixels in the shared reference/candidate image space. Both images must have identical dimensions and the rectangle must be fully inside them. Omitting the option preserves full-frame behavior.
+
+The existing 15 WMO-ready Goldshire candidates were rescored inside `[300,100,1400,650]`, covering the reference facade/chimney and nearby trees. ROI error ranked FOV 38/yaw 10/pitch -15 first at `0.081777` with changed-pixel ratio `0.697185`. This is an interior pitch candidate and improves on the full-frame grass preference, but human inspection still rejects manifest calibration: the facade is too large, its upper part is cropped, and the rectangle retains substantial road/terrain pixels. A common rectangle is useful diagnostic weighting, not semantic landmark registration. Explicit approved landmark coordinates or a mask remain necessary.
+
 During this run the sweep orchestration was hardened in two ways: expected comparator exit code `1` is now collected through an explicit child process, and viewport dimensions are explicit. A reference/candidate `size_mismatch` now fails before ranking instead of producing an empty metric.

docs/M00_FINAL_PAIRED_EVIDENCE.md

@@ -0,0 +1,60 @@
+# M00 Final Paired Evidence — 2026-07-12
+
+## Scope and result
+
+The final local M00 audit used Godot 4.6.1, the existing cache state, the manifest-standard 8-second readiness wait, and `2560x1440` visual candidates. Five private WoW 3.3.5a build 12340 references produced ten cold/warm comparisons with no missing pair. Private images and generated reports remain outside Git.
+
+The command completed even though every pair exceeded the strict regression tolerance. Comparator exit code `1` means a valid report with visual gaps; only codes outside `0/1` are infrastructure failures.
+
+```powershell
+.\tools\run_render_baseline.ps1 `
+  -Output user://render_baseline_m00_final_2026-07-12 `
+  -CacheState existing `
+  -WaitSeconds 8 `
+  -MeasureSeconds 0.5 `
+  -ReferenceCheckpointDirectory .\sources\OpenWCReferenceCheckpoints `
+  -VisualComparisonReport user://render_baseline_m00_final_2026-07-12/visual_comparison.json `
+  -VisualComparisonWidth 2560 `
+  -VisualComparisonHeight 1440
+```
+
+## Paired metrics
+
+| Checkpoint | Pass | Mean perceptual error | Changed-pixel ratio | Classification |
+|---|---|---:|---:|---|
+| `elwynn_adt_boundary` | cold | 0.087347 | 0.634984 | Composition plus terrain/vegetation/material gaps |
+| `elwynn_adt_boundary` | warm | 0.087255 | 0.634598 | Stable against revisit |
+| `elwynn_terrain_overview` | cold | 0.174136 | 0.851806 | Largest composition/framing gap |
+| `elwynn_terrain_overview` | warm | 0.176952 | 0.861275 | Stable gap, not a cold-load artifact |
+| `elwynn_waterfall_liquid` | cold | 0.088526 | 0.706897 | Framing, terrain, liquid and lighting gaps |
+| `elwynn_waterfall_liquid` | warm | 0.087490 | 0.695281 | Stable against revisit |
+| `goldshire_dense_m2` | cold | 0.066892 | 0.575205 | Smallest available paired gap; vegetation/material differences remain |
+| `goldshire_dense_m2` | warm | 0.066927 | 0.575547 | Stable against revisit |
+| `goldshire_inn_large_wmo` | cold | 0.101230 | 0.701836 | WMO is ready; manual framing and presentation gaps remain |
+| `goldshire_inn_large_wmo` | warm | 0.101203 | 0.701855 | Stable against revisit |
+
+Cold and warm values are close for all five references. The paired gaps are therefore not explained by first-visit cache readiness. Goldshire timing evidence separately demonstrated that the 8-second wait is necessary for WMO publication.
+
+## Fidelity interpretation
+
+M00 now has reproducible quantitative evidence for terrain, ADT boundary, dense M2, large WMO and liquid checkpoints. The reports demonstrate current differences; they do not claim renderer parity. Camera calibration experiments also show that unknown manual yaw/pitch/FOV cannot be recovered reliably from full-frame color error alone. Region scoring reduces terrain bias but still requires semantic landmark approval.
+
+Two manifest checkpoints still lack original-client paired references:
+
+- `gryphon_roost_native_animation` is a synthetic native-animation probe and has no matching build 12340 fixture.
+- `elwynn_sky_dusk` has no approved build 12340 sky-transition fixture at the same time/weather state.
+
+These must either receive private reference evidence or be explicitly accepted by the milestone integrator as unpaired diagnostic coverage. Until that decision, the target remains `ACTIVE`.
+
+## Operational risks
+
+The full GUI capture completed and wrote both reports, but Godot still emitted renderer RID/resource leak diagnostics at exit. The earlier anonymous headless `RefCounted` leak is fixed; GUI renderer allocations are a distinct known risk. They do not invalidate the written images or metrics, but must remain recorded rather than hidden.
+
+## Integrator decision checklist
+
+- Five available reference checkpoints: paired cold/warm report complete, missing pairs zero.
+- Animated M2 fixture: missing or explicitly accepted as synthetic-only.
+- Sky-transition fixture: missing or explicitly accepted as unpaired.
+- Renderer visual differences: registered as gaps, not parity claims.
+- GUI renderer RID/resource diagnostics: accepted as a tracked risk or assigned a follow-up blocker.
+- Target status and `OPENWC_TARGET_DONE` marker: integrator-only action.

docs/modules/world-renderer.md

@@ -232,6 +232,9 @@ Exact exported settings and cache versions remain documented in [`../../RENDER.m
 - Camera-pose sweep can now rank bounded yaw/pitch grids without changing manifest defaults; perceptual ranking remains diagnostic and requires human framing approval.
 - The first Goldshire pose grid was invalidated by WMO readiness: the inn is absent after 2 seconds but visible after the manifest-standard 8-second wait. Pose comparisons must stabilize asynchronous checkpoint content first.
 - A ready-scene Goldshire 3x3 grid improved mean error from `0.101402` at zero offsets to `0.087952` at yaw `10`/pitch `-10`; pitch/FOV refinement remains required before manifest calibration.
+- Joint Goldshire pitch/FOV refinement had no human-approved optimum: full-frame error preferred grass-heavy FOV 62/pitch -25, while FOV 38/pitch -10 better matched building scale. Landmark/region scoring is required before calibration.
+- Common-region scoring `[300,100,1400,650]` moved Goldshire's numeric optimum to FOV 38/yaw 10/pitch -15, but human inspection still found an oversized/cropped facade; semantic landmarks or masks remain required.
+- Final paired audit compared ten cold/warm images for five private build 12340 references with zero missing pairs; mean error ranges `0.066892..0.176952`. Animated-M2 and sky-transition references remain an integrator acceptance decision.
 - D3D12 descriptor issues remain; the capture-path anonymous `RefCounted` shutdown leak is regression-covered by a clean verbose dry-run, while other RID/resource diagnostics still require independent evidence.
 - M2/WMO/material/particle/ribbon/portal parity incomplete.
 - Public API is mostly exported configuration rather than stable contracts.
@@ -261,5 +264,6 @@ Exact exported settings and cache versions remain documented in [`../../RENDER.m
 - [`../../RENDER.md`](../../RENDER.md)
 - [`../RENDER_BASELINE.md`](../RENDER_BASELINE.md)
 - [`../CAMERA_POSE_SWEEP.md`](../CAMERA_POSE_SWEEP.md)
+- [`../M00_FINAL_PAIRED_EVIDENCE.md`](../M00_FINAL_PAIRED_EVIDENCE.md)
 - [`../../targets/roadmap/02-rendering-and-graphics.md`](../../targets/roadmap/02-rendering-and-graphics.md)
 - [`../GODOT_BEST_PRACTICES.md`](../GODOT_BEST_PRACTICES.md)

src/tools/compare_render_checkpoints.gd

@@ -3,6 +3,7 @@ extends SceneTree
 ## Compares original-client JPG/PNG references with OpenWC checkpoint PNGs.
 ## Usage: godot --headless --path . --script res://src/tools/compare_render_checkpoints.gd --
 ##   --reference <directory> --candidate <directory> [--output <report.json>]
+##   [--region x,y,width,height]
 ##   [--pixel-threshold 0.04] [--mean-threshold 0.015] [--changed-ratio-threshold 0.01]
 ## Use --self-test for a synthetic regression check that writes only under user://.
 
@@ -56,13 +57,18 @@ func _parse_arguments(raw_arguments: PackedStringArray) -> Dictionary:
 			parsed.self_test = true
 			index += 1
 			continue
-		if argument not in ["--reference", "--candidate", "--output", "--only", "--pixel-threshold", "--mean-threshold", "--changed-ratio-threshold"]:
+		if argument not in ["--reference", "--candidate", "--output", "--only", "--region", "--pixel-threshold", "--mean-threshold", "--changed-ratio-threshold"]:
 			return {"error": "unknown argument: %s" % argument}
 		if index + 1 >= raw_arguments.size():
 			return {"error": "missing value for %s" % argument}
 		var key := argument.trim_prefix("--").replace("-", "_")
 		var value := raw_arguments[index + 1]
-		if key.ends_with("threshold"):
+		if key == "region":
+			var region_parse_result := _parse_comparison_region(value)
+			if region_parse_result.has("error"):
+				return region_parse_result
+			parsed[key] = region_parse_result.region
+		elif key.ends_with("threshold"):
 			if not value.is_valid_float():
 				return {"error": "%s requires a number" % argument}
 			parsed[key] = float(value)
@@ -72,6 +78,24 @@ func _parse_arguments(raw_arguments: PackedStringArray) -> Dictionary:
 	return parsed
 
 
+func _parse_comparison_region(value: String) -> Dictionary:
+	var components := value.split(",", false)
+	if components.size() != 4:
+		return {"error": "--region requires x,y,width,height"}
+	for component in components:
+		if not component.strip_edges().is_valid_int():
+			return {"error": "--region components must be integers"}
+	var region := Rect2i(
+		int(components[0]),
+		int(components[1]),
+		int(components[2]),
+		int(components[3])
+	)
+	if region.position.x < 0 or region.position.y < 0 or region.size.x <= 0 or region.size.y <= 0:
+		return {"error": "--region requires non-negative x/y and positive width/height"}
+	return {"region": region}
+
+
 func _compare_directories(reference_directory: String, candidate_directory: String, options: Dictionary) -> Dictionary:
 	reference_directory = ProjectSettings.globalize_path(reference_directory)
 	candidate_directory = ProjectSettings.globalize_path(candidate_directory)
@@ -108,6 +132,7 @@ func _compare_directories(reference_directory: String, candidate_directory: Stri
 		"pixel_threshold": options.pixel_threshold,
 		"mean_threshold": options.mean_threshold,
 		"changed_ratio_threshold": options.changed_ratio_threshold,
+		"comparison_region": _region_array(options.get("region", Rect2i())),
 		"compared_count": results.size() - missing_count,
 		"failed_count": failed_count,
 		"missing_count": missing_count,
@@ -127,12 +152,22 @@ func _compare_images(reference_path: String, candidate_path: String, options: Di
 			"reference_size": [reference_image.get_width(), reference_image.get_height()],
 			"candidate_size": [candidate_image.get_width(), candidate_image.get_height()],
 		}
+	var comparison_region: Rect2i = options.get("region", Rect2i(Vector2i.ZERO, reference_image.get_size()))
+	if comparison_region == Rect2i():
+		comparison_region = Rect2i(Vector2i.ZERO, reference_image.get_size())
+	var image_region := Rect2i(Vector2i.ZERO, reference_image.get_size())
+	if not image_region.encloses(comparison_region):
+		return {
+			"status": "region_out_of_bounds",
+			"region": _region_array(comparison_region),
+			"image_size": [reference_image.get_width(), reference_image.get_height()],
+		}
 
 	var error_sum := 0.0
 	var changed_pixel_count := 0
-	var pixel_count := reference_image.get_width() * reference_image.get_height()
+	var pixel_count := comparison_region.size.x * comparison_region.size.y
-	for y in reference_image.get_height():
+	for y in range(comparison_region.position.y, comparison_region.end.y):
-		for x in reference_image.get_width():
+		for x in range(comparison_region.position.x, comparison_region.end.x):
 			var perceptual_error := _perceptual_color_error(reference_image.get_pixel(x, y), candidate_image.get_pixel(x, y))
 			error_sum += perceptual_error
 			if perceptual_error > float(options.pixel_threshold):
@@ -144,11 +179,18 @@ func _compare_images(reference_path: String, candidate_path: String, options: Di
 		"status": "passed" if passed else "threshold_exceeded",
 		"width": reference_image.get_width(),
 		"height": reference_image.get_height(),
+		"region": _region_array(comparison_region),
 		"mean_perceptual_error": mean_perceptual_error,
 		"changed_pixel_ratio": changed_pixel_ratio,
 	}
 
 
+func _region_array(region: Rect2i) -> Array[int]:
+	if region == Rect2i():
+		return []
+	return [region.position.x, region.position.y, region.size.x, region.size.y]
+
+
 func _perceptual_color_error(reference_color: Color, candidate_color: Color) -> float:
 	var reference_linear := reference_color.srgb_to_linear()
 	var candidate_linear := candidate_color.srgb_to_linear()
@@ -229,6 +271,22 @@ func _run_self_test() -> int:
 	if not identical_report.passed or changed_report.passed:
 		push_error("RENDER_CHECKPOINT_DIFF SELF_TEST: expected identical pass and changed failure")
 		return 1
+	var region_options := options.duplicate()
+	region_options["region"] = Rect2i(1, 1, 1, 1)
+	var excluded_change_report := _compare_directories(reference_directory, changed_directory, region_options)
+	if not excluded_change_report.passed or excluded_change_report.comparison_region != [1, 1, 1, 1]:
+		push_error("RENDER_CHECKPOINT_DIFF SELF_TEST: region must exclude pixels outside the rectangle")
+		return 1
+	var out_of_bounds_options := options.duplicate()
+	out_of_bounds_options["region"] = Rect2i(1, 1, 2, 2)
+	var out_of_bounds_result := _compare_images(
+		reference_directory.path_join("synthetic.jpg"),
+		changed_directory.path_join("synthetic__cold_process.png"),
+		out_of_bounds_options
+	)
+	if out_of_bounds_result.status != "region_out_of_bounds":
+		push_error("RENDER_CHECKPOINT_DIFF SELF_TEST: out-of-bounds region must fail")
+		return 1
 	if reference_image.save_jpg(reference_directory.path_join("ignored.jpg"), 1.0) != OK:
 		return 1
 	var filtered_options := options.duplicate()

tools/run_render_baseline.ps1

@@ -40,6 +40,15 @@ function Invoke-GodotStep {
     }
 }
 
+function Invoke-GodotComparisonStep {
+    param([string]$Name, [string[]]$Arguments)
+    Write-Host "[$Name] $GodotPath $($Arguments -join ' ')"
+    $process = Start-Process -FilePath $GodotPath -ArgumentList $Arguments -NoNewWindow -Wait -PassThru
+    if ($process.ExitCode -notin @(0, 1)) {
+        throw "$Name failed with exit code $($process.ExitCode)"
+    }
+}
+
 Push-Location $repoRoot
 try {
     Invoke-GodotStep "project-load" @("--headless", "--path", ".", "--quit")
@@ -80,7 +89,7 @@ try {
 			"--viewport-height", $VisualComparisonHeight.ToString()
 		)
 		Invoke-GodotStep "visual-capture" $visualCaptureArgs
-        Invoke-GodotStep "visual-comparison" @(
+        Invoke-GodotComparisonStep "visual-comparison" @(
             "--headless", "--path", ".",
             "--script", "res://src/tools/compare_render_checkpoints.gd",
             "--",

tools/sweep_render_checkpoint_camera_pose.ps1

@@ -9,8 +9,10 @@ param(
     [double[]]$YawOffsets = @(-10.0, 0.0, 10.0),
     [double[]]$PitchOffsets = @(-10.0, 0.0, 10.0),
     [double]$CameraFov = 62.0,
+    [double[]]$CameraFovValues = @(),
     [int]$ViewportWidth = 1280,
     [int]$ViewportHeight = 900,
+    [int[]]$ComparisonRegion = @(),
     [double]$WaitSeconds = 8.0,
     [double]$MeasureSeconds = 0.1,
     [switch]$PlanOnly
@@ -32,9 +34,13 @@ if (-not (Test-Path -LiteralPath $GodotPath)) {
 if (-not (Test-Path -LiteralPath $ReferenceDirectory)) {
     throw "Reference directory not found: $ReferenceDirectory"
 }
+if ($ComparisonRegion.Count -notin @(0, 4)) {
+    throw "ComparisonRegion requires x,y,width,height"
+}
 
 $invariantCulture = [Globalization.CultureInfo]::InvariantCulture
-$cameraFovText = $CameraFov.ToString($invariantCulture)
+$isJointFovSweep = $CameraFovValues.Count -gt 0
+$effectiveCameraFovValues = if ($isJointFovSweep) { $CameraFovValues } else { @($CameraFov) }
 $waitSecondsText = $WaitSeconds.ToString($invariantCulture)
 $measureSecondsText = $MeasureSeconds.ToString($invariantCulture)
 $absoluteOutput = if ($Output.StartsWith("user://")) {
@@ -46,11 +52,16 @@ $ranking = [Collections.Generic.List[object]]::new()
 
 Push-Location $repoRoot
 try {
-    foreach ($yawOffset in $YawOffsets) {
+    foreach ($cameraFovValue in $effectiveCameraFovValues) {
-        foreach ($pitchOffset in $PitchOffsets) {
+        foreach ($yawOffset in $YawOffsets) {
+            foreach ($pitchOffset in $PitchOffsets) {
+            $cameraFovText = $cameraFovValue.ToString("0.###", $invariantCulture)
             $yawText = $yawOffset.ToString("0.###", $invariantCulture)
             $pitchText = $pitchOffset.ToString("0.###", $invariantCulture)
             $candidateName = "yaw_$($yawText.Replace('-', 'n').Replace('.', '_'))__pitch_$($pitchText.Replace('-', 'n').Replace('.', '_'))"
+            if ($isJointFovSweep) {
+                $candidateName = "fov_$($cameraFovText.Replace('.', '_'))__$candidateName"
+            }
             $candidateOutput = "$Output/$candidateName"
             $comparisonReport = Join-Path $absoluteOutput "$candidateName.json"
 
@@ -58,7 +69,7 @@ try {
                 $ranking.Add([pscustomobject]@{
                     yaw_offset_degrees = $yawOffset
                     pitch_offset_degrees = $pitchOffset
-                    camera_fov_degrees = $CameraFov
+                    camera_fov_degrees = $cameraFovValue
                     candidate_directory = $candidateOutput
                 })
                 continue
@@ -82,6 +93,9 @@ try {
                 "--reference", $ReferenceDirectory, "--candidate", $candidateOutput,
                 "--only", $Checkpoint, "--output", $comparisonReport
             )
+            if ($ComparisonRegion.Count -eq 4) {
+                $comparisonArguments += @("--region", ($ComparisonRegion -join ","))
+            }
             $comparisonProcess = Start-Process -FilePath $GodotPath -ArgumentList $comparisonArguments -Wait -PassThru -NoNewWindow
             if ($comparisonProcess.ExitCode -notin @(0, 1)) {
                 throw "Comparison failed for yaw=$yawText pitch=$pitchText"
@@ -101,12 +115,13 @@ try {
             $ranking.Add([pscustomobject]@{
                 yaw_offset_degrees = $yawOffset
                 pitch_offset_degrees = $pitchOffset
-                camera_fov_degrees = $CameraFov
+                camera_fov_degrees = $cameraFovValue
                 mean_perceptual_error = $meanError
                 changed_pixel_ratio = $changedRatio
                 candidate_directory = $candidateOutput
                 comparison_report = $comparisonReport
             })
+            }
         }
     }
 
@@ -123,11 +138,12 @@ try {
         viewport = @($ViewportWidth, $ViewportHeight)
         wait_seconds = $WaitSeconds
         measure_seconds = $MeasureSeconds
+        comparison_region = @($ComparisonRegion)
         candidates = $rankedCandidates
     }
     New-Item -ItemType Directory -Force -Path $absoluteOutput | Out-Null
     $rankingDocument | ConvertTo-Json -Depth 6 | Set-Content -Encoding UTF8 -LiteralPath $rankingPath
-    $rankedCandidates | Format-Table yaw_offset_degrees, pitch_offset_degrees, mean_perceptual_error, changed_pixel_ratio
+    $rankedCandidates | Format-Table camera_fov_degrees, yaw_offset_degrees, pitch_offset_degrees, mean_perceptual_error, changed_pixel_ratio
     $completionKind = if ($PlanOnly) { "plan" } else { "ranking" }
     Write-Output "Camera pose sweep $completionKind completed: $rankingPath"
 } finally {