Python tool for aircraft Doppler audio geolocation using historical ADS-B flight data (Source-available)
Integration verification, 2026-11-03 UTC
The primary integration agent reviewed and combined the worker branches in an isolated worktree, then ran the repository suite against the actual combined source in a clean Python 3.14 environment. No system site packages were enabled.
| Combined increment | Tests | Runtime | Tested commit |
|---|---|---|---|
| Evaluation, discovery, repaired index, independent index regressions, partial association | 94 passed | 113.85 s | 5e40a563956a304c0acecfa5b39a322a0a93197e |
| Fragment discovery, national tiled driver, full perturbation profiles | 104 passed | 57.49 s | afec84d2526c1eeaf016b94e466bc4b510908e37 |
| Per-flight altitude profiles, codec/replay verifiers, benchmark-clock regressions | 121 passed | 62.24 s | 1d74fcd880ce5a9873b5661428575e2954064b76 |
| All-event matcher, canonical renderer, refinement-grid union, confidence/evaluation and preserved real/synthetic evidence | 175 passed | 66.39 s | e3d1952b497e16d3d5066e6bbbe5ff05c8e0310f |
| All-event tiled driver, trusted-window edges, acquisition controls and raw-motion audit | 218 passed + 6 subtests | 51.94 s | c91de8214f4064ee5be4c2a24f29802949a9e812 |
| Main-based bounded chunk reader and Linux live-RSS regression | 219 passed + 6 subtests | 51.61 s | 848c190699d17fee6f7528e310b457ecc973cfea |
The command was PYTHONPATH=src <clean-venv>/bin/python -m pytest -q --junitxml=<receipt>. The final run also set OPENBLAS_NUM_THREADS=1 and
OMP_NUM_THREADS=1. Earlier runs emitted one pandas future-dtype warning from the index
aircraft-class fallback; W2’s explicit object conversion removes it in the218-test
increment. There were no failures or skips. Machine-readable
receipts and JUnit reports
record the exact source objects and environment. The 175-test receipt
and JUnit XML preserve the latest frozen combination. The published W4 follow-up
changes fixture altitude metadata/documentation; production source and tests are
identical to that tested combination. Test counts cover integration
behavior; they do not establish real-world location accuracy.
The user's updated acceptance target is at most 1,000 m error, with 100–500 m or better preferred. Future evaluations report 100/500/1,000 m thresholds and must exercise three, four and more independent flyovers without a fixed event cap. Older runs retain their original settings and thresholds; changing the target does not repair missing reference coordinates, clock evidence or flight coverage.
Current real-data evidence is deliberately narrower:
- The clean configured replay preserves both the regional two-event success and the faint-event result outside the 100 m target.
- The metadata audit does not establish independent recorder coordinates or sample-zero UTC for the separate 40-passage dataset.
- The automatic synthetic execution and context diagnostic retains the inconclusive result and its competing associations. The extra spectral-context experiment is development evidence, not a calibrated decision rule or a replacement for the frozen run.
- Corrected Times Square and Liberty recordings each decode to 300 seconds and yield no proposals under unchanged discovery. Neither establishes microphone sample-zero UTC or surveyed coordinates, so neither is eligible for location-error acceptance at any threshold.
- Canonical independent 3/4/6-flyover fixtures produced 13/13 detected events with frozen settings. Their code, complete discovery evidence and checksummed media are published. Flight association, multi-event location acceptance and the finite-search alternative assessment still require the subsequent actual product replays.
The source YouTube MP4 is excluded. Earlier experiments and failed outcomes have not been overwritten. See the coordination board for active integration dependencies and worker replies.
The main-based driver receipt preserves the combined all-event increment and native JUnit. Its sole merge conflict was independent appended tiled tests, both retained. This source uses all prepared groups by default, keeps explicitly capped work partial, and preserves the unsupported-motion/coverage limits measured in the raw audit. Product integration and independent real acceptance remain separate gates.
A separate product-review combination passes254 tests plus14 subtests at401b2e7. It includes actual service and independent decision controls, but remains a separate source from main and does not certify an accepted location. The main chunk-memory receipt records219 plus6, with no skips. W1’s window-reader experiment uses live process RSS and exact canonical six-event equivalence. This fixes mapped whole-file residency; it does not identify the cause of the earlier1.6GiB outlier or bound total search/browser memory.