[Whyater]

Overnight work on making the sim usable with real measured inputs. No registered result folder or hash in tools/reproduce_registry.json was changed. Most of this is new side-run tools, docs and tests; existing input tools also gain guards that can now refuse or relabel an input (listed below).

What changed

Bench data and measured-input chain

• tools/preflight_bench_data.py: checks a bench folder before any intake runs.
• tools/bench_verdict.py: one command from preflight through intakes, combined rerun and latency budget.
• tools/propose_measured_entry.py: drafts a registry entry for a measured run, with strict run-dir checks.
• tools/measured_provenance.py and edits to the link, combined and measured reruns, latency budget and fit writers: a "measured" label now needs provenance, otherwise it becomes "unverified".
• Camera log quirk checks in sim/vision_noise_fit.py, sim/detection_log.py and tools/intake_vision.py, documented in docs/DETECTION_LOG.md.
• Combined S6 rerun gains default-off noise shape, bias and dropout options.
• Input guards: tools/latency_budget.py refuses a typed --processing-median-ms outside 0.5 to 500 ms and reports "not measured" when no input is measured; tools/vision_latency.py checks --refresh-hz against the recorded display_refresh_hz; sim/vision_noise_fit.py refuses positions that are not in meters and leaves latency unfitted when any frame is over 5 s.

Radio

• sim/radio/packet.py: fixed 161-byte ESP-NOW state packet with version byte and CRC-32, under the 250-byte v1.0 limit.
• tools/espnow_airtime_budget.py: airtime and channel use per swarm size and send rate.

Side runs (all simulated; outputs stay out of sim/runs/)

• Grid extension past the old delay and loss ceilings, and follow-ups.
• Outlier robustness with an NIS gate, turn-aware CT and IMM trackers, switching targets, and gated constant acceleration.
• Vision error shape and a camera-shaped bias model.
• Slow-motor gains, drag long check, send-period hold correction, monitor mission noise sweep.
• Late-packet re-filtering at 2x and 3x jitter, with an independent capture-order check.
• Full-chain closed-loop pilot, plus an opt-in lead goal rule (--lead: a constant-turn prediction aimed the 0.75 s controller lag plus the setpoint radio delay ahead; the old velocity rule stays the default and its rows are unchanged) and a pure delay sweep (--delay-sweep). One hand check (circle radius) failed against the sim and is kept as a strict xfail with its reason.
• Tracker float32 and fixed-tick numeric readiness, plus runs explaining why snapping a capture to the next 100 Hz tick looked about 1 % better than fusing at the exact time: it comes from scoring at the 10 ms ticks (on a 1 ms scoring grid the sign flips) and from a filter q larger than a straight line needs. No bug in the exact-time path; exact-time fusion stays the reference.

Platform and test hygiene

• LF line endings for result writers so Windows reruns match registered hashes.
• Off-macOS float tolerance in the negative-control comparisons of the vision error shape and grid-extension side runs.
• Platform tolerance test with planted bugs; corrected comms hand checks (four stale xfails removed).
• tools/tests/test_reproduce_registry_coverage.py: byte-checks every committed registered output and fails if a new unregistered result folder appears. Six existing unregistered folders are allowlisted for review (burst-bridge-radio, d4-ellipsoid-diagnosis, e-exp035-logs, missions-atlas, retry-fades-exp038, s6-interaction).
• Opt-in ground plane (default off).

Testing

• Each merge was tested on Windows with its own test files; the measured-input chain (10 files) passed 373 tests on the merged branch.
• CI (windows-2022, full suite) passed on 547d340 (both runs). On 304e929 the push run passed and the pull-request run failed one exact-hash check in sim/tests/test_ground.py: the same code gave different bits on a runner with a different CPU. That check now compares within a relative 1e-9 (09ce9e0). The run on the current head is in the checks below.
• A final whole-PR review read every file touched by more than one merge and found no broken logic; it asked for description fixes and a naming cleanup. A re-review found four more internal labels in side-run tools and the ground test; those are replaced in ae3d33d and 3d55938 (the final ground test profile name gives a derived profile identical to the old one in every field).

Not in this PR

• Echo firmware for the radio bench (not started).
• Branches that still need a decision: radio log quirks, frame and time layer, public-number linter, camera bias scorer, radio airtime model fixes.