Stage 0: six gaps a second site exposed, fixed one commit each

[PSheon]

What this fixes

Six gaps in Stage 0 that a second site exposed (FTI factory, 2026-09-10/11; record in /home/paul/hydranet_fti_stage0/runs/REVIEW.md). One commit each; every commit stands alone and carries its test.

commit	gap	what changed
fix(demo_video)	track boxes drawn with / 2.0 — a 1920-wide source as a constant; 2/3 off on 1280, 2× on 3840	analytics.world.panel_box: the panel/source ratio per axis, used by the demo
fix(commissioning)	pick_daytime_slot hardcoded UTC+8; a US-Central site passed the daytime gate by coincidence	cameras.json carries utc_offset_hours (8 when absent); static_plates records it; onboard_camera passes it
feat(onboard)	one camera pinned by name; floor_calibrate's measured vfov had no consumer	--vfov-pins reads the instrument's pins file over BUILTIN_PINS; the pin is the camera's primary row and joins the sweep; vfov_source/k1_source say where each number came from
fix(rulers)	catalogue consensus fooled by shared depth bias (600 mm floor read 0.85/0.74 → picked 0.80); person_ruler anchored factor 1 to a bootstrap scale	Ruler.anchored; person_ruler is None on unmeasured/bootstrap; combine returns an explicit unanchored verdict instead of "within 10%"; scale_rulers marks the consensus tile unanchored
feat(tools)	30 tools derived ROOT from their own file — a second site meant a second copy of the code	every tool honours SYNCAI_ROOT (the pattern scene_mesh.py already used); shipped.py resolves checkpoints through paths.repo_root; a test keeps bare Path(__file__) roots out of the chain
feat(camera_json)	contract had no native frame size; a 720p or 4K stream returned metres, not an error	CameraFile.source_size_px (optional; v3 files read unchanged), recorded by static_plates, carried by onboard_camera, kept by from_onboard_calib, aspect-checked in validate(), and demo_video says when the clip is not the commissioned stream
docs(figures)	the figure tax	both README GIFs re-cut last with the recorded render arguments — byte-identical to the published ones; only the audit sidecars (render, commit stamp) move

CI-equivalent locally: ruff format --check, ruff check, scripts/coverage_ratchet.sh — 3581 passed, 16 skipped; the two figure-staleness failures were cleared by the last commit. One local-only failure is not this branch: test_map_anything_eval asserts the 8 shipped cameras and this box now holds 9 (Tao-Hsin-cam15, another session) plus test1; CI has no runs/ and skips it.

Deliberately not in this PR

• Files not yet in git (another session's in-flight Gate D work: commission_camera.py, floor_calibrate.py, geometry_cache.py, rebuild_geometry.py, ground_control.py): the vfov instrument's no-crossing fallback (cam632's 48→50° sign flip is a plane-pick flip reported as "no crossing"), a labelled bootstrap route for personless cameras in commission_camera, and depth_scale in recipe-built caches (scale_rulers.apply refuses them) all land there. Listed in REVIEW.md with evidence; to follow once that work is committed.
• Site name inside the paths (datasets/studioa_static, runs/commission01): root is now switchable, the literal names are not — a paths config is the next step, after the dirty tools land.
• GeomTeacher's runs/zones01 dependency (it only uses the walkable polygon propose_zones derives): same file as above.
• Design decisions the code cannot make: a rack/machine/pallet class for the structure vote (racks are wall today), absolute object rulers beyond the tile catalogue, cross-camera registration into one world, and thresholds on overlay IoU (Gate D, in flight elsewhere).

Evidence the changes act on

• FTI-SMT-cam815 (1280×720): boxes at 2/3 of the people before, on them after (demo_video, check frame 112).
• FTI-MDF-cam821/cam632 (3840×2160): the catalogue picked 0.80 m for a 600 mm raised floor; the folding table (0.74 m) settled it at ×0.845 vs the tile's ×0.847.
• FTI-MDF-cam821: floor_calibrate measured 62.8°; applied by hand before, --vfov-pins now.

🤖 Generated with Claude Code