Witnesses for the clipped coverage grid and fitted ground

[obvious-autobuild]

What this adds

Pipeline-level witnesses on top of ea38351 (Coverage stops at the fitted extent and samples the fitted ground), which fixed the two coverage defects in the handoff (issues #95 and #96). No behavior change here — tests, a synthetic bundle builder, a schema-compat check script, and a results record.

• recon/tests/witness_bundle.py — renders synthetic scan bundles from analytic planes (wall pieces, a back wall, a hole, a falling floor) with exact poses plus fixed-seed noise, into the app's bundle layout (scene.json, JPEGs, Float32 depth, UInt8 confidence). Nothing synthetic claims to be a real capture: no home, no private rule, no real-world coordinate.
• recon/tests/test_coverage_pipeline.py — eight witnesses through pipeline.run, the caller a real bundle takes, with --depth lidar (no model, no network). Oracle is the written scene.json / coverage.json / geometry.json, never a test-local copy of the clipping logic: exported spans stop at the fitted extent while the phone's baseline claims past it; a hole in the wall stays unobserved across a bridged fit; ground follows a falling fitted plane (0.35 m/m) where the old horizontal-plane sampler cannot produce a 2 ft entry at all; short (0.6 m) and empty wall tracks refuse; no-keyframes, single-LiDAR-frame and too-close-tracks refuse or observe nothing.
• recon/tests/scene_schema_check.py — validates an exported scene.json against server/schemas/scene.schema.json; reports the input bundle's per-keyframe depth passthrough (an app-side extension the schema does not define) on its own line rather than folding it into a pass.
• recon/results/coverage_extent_witnesses.md — baseline counts, witness table, owned smoke command with output hashes, and a field-level old-vs-new comparison run at c813182.

Check commands (all run in recon/, offline)

Command	Result
uv sync --locked	resolves cleanly, 17 packages
uv run ruff check .	All checks passed
uv run ruff format --check .	clean
uv run pytest -q	98 passed, 56 warnings
uv run pytest -q tests/test_coverage_pipeline.py	8 passed

Owned CLI smoke (exact command, recorded in the results doc with output hashes):

uv run python tests/witness_bundle.py /tmp/witness-smoke/bundle
uv run python -m recon /tmp/witness-smoke/bundle --out /tmp/witness-smoke/out --work /tmp/witness-smoke/work --depth lidar --no-server

Old-vs-new on the same bundle (c813182 vs head): grid now starts at the fitted extent (-3.035, was -3.5), cells_end_ft appears and clips the boundary cell, wall span covers the full fit ([−3.0, 5.332], was [−3.0, 5.0]), overhead stops claiming past the fit ([-3.035, 5.332], was [-3.5, 5.5]), phantom last grid cell goes false → true-observed-at-real-end.

Not run

• The depth/model path (--depth moge2 or photo-based depth): these bundles are LiDAR-rendered by design — running the model would need its process and weights.
• The server round-trip and server/tests: recon-side only; the schema-compat receipt covers the exported document against the committed schema.
• ETH3D acceptance (make accept) and the iOS checks: untouched by this branch, not rerun.
• Any run against a real home capture.

Assumptions

• The witnesses treat server/schemas/scene.schema.json as the compatibility bar for the exported scene.json; the input depth extension is reported separately, not hidden.
• Free space keeping its own weaker bar than the two-position rule is pinned as current behavior (commented in the test), not fixed here — it is outside this pass's scope.
• recon/results/ is the right home for the witness record, next to the existing ETH3D results.

Handoff (from the recon handoff, done / remaining)

• Done here: pipeline witnesses for issues Sample reconstruction ground coverage on the fitted plane #95/Keep reconstruction coverage intervals inside fitted wall extents #96 incl. holes, sloping ground, short/empty tracks and clipped corner cases; schema-compat receipts; smoke hashes and an old-vs-new coverage comparison.
• Remaining: depth/model (MoGe-2) path coverage on synthetic images, server-side acceptance, iOS build checks, and any run on real captures — all need paid services or environments not available offline.

Review request: the repair under test touches the recon coverage area — @hunter, this will need your review.

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 3d1f2695-aa80-4f44-8714-0df8b95b499c

---

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}