Assess Denver, and freeze the Stage 1 inventories for replication (#96, #59)

.gitignore

@@ -36,11 +36,61 @@ analysis_out/op/*
 !analysis_out/op/*.csv
 !analysis_out/op/*.json
 
+# Stage 1's coordinate-tolerance result (#96): a few KB, and it is what converts an
+# offset distribution into a decision. Re-derivable on CPU from the committed
+# benchmark bundles plus any verdicts file.
+!analysis_out/stage1_offset_tolerance.json
+# Per-city tolerance runs (--verdicts <city>), e.g. charlotte-nc. Same few KB, and
+# they are what turn a city's offset distribution into a labels-lost number.
+!analysis_out/stage1_offset_tolerance_*.json
+
+# Stage 1's realised bearing residual (#96): a few KB, and it is the null that makes
+# any future city's registration number interpretable. Regenerating it needs a Hub
+# fetch, so committing it keeps the 5j table checkable from a clean checkout.
+!analysis_out/stage1_bearing_residual.json
+
 # E1's result (#59): a few KB, and committing it makes the label-ceiling numbers
 # checkable without re-running the Hub fetch the script needs to regenerate them.
 !analysis_out/e1_stage1_label_recall.json
 !analysis_out/miss_decomposition.json
 
+# Per-ramp/per-corner geometry reports (#96): a few KB each, and they hold the exact
+# numbers quoted in docs/curb_ramp_data_sourcing.md 5d, re-derivable on CPU from the
+# committed snapshots under data/inventories/.
+!analysis_out/inventory_geometry_*.json
+
+# Location-precision review sheets (#96). The verdicts -- and the tile keys recording
+# which imagery produced them -- are the irreplaceable part, so verdicts.json is tracked.
+# review_sheet.html is not: it embeds basemap tiles under terms that do not permit
+# re-hosting, and it regenerates from the committed script plus the recorded seed. The
+# tile caches beside it are likewise someone else's imagery.
+!analysis_out/review_*/
+analysis_out/review_*/*
+!analysis_out/review_*/verdicts.json
+# summary.json too: it is the reduction quoted in 5f, a few KB, and committing it
+# means a reader can check a figure in the prose against the table it came from
+# without re-running anything. Regenerable from verdicts.json by
+# scripts/analysis/inventory_review_summary.py.
+!analysis_out/review_*/summary.json
+
+# Georeference check (#96): the JSON is the evidence that the sheet's crosshair and
+# rings are true, and the two PNGs are what a reader looks at to agree. Small, and
+# the whole point is that a claim of correct registration is checkable rather than
+# asserted. The tile cache beside them is someone else's imagery.
+# One directory per city since §5i (georef_check_denver-co, georef_check_seattle-wa),
+# plus the _chips variant that measures under a specific review's own verdicts.
+!analysis_out/georef_check_*/
+analysis_out/georef_check_*/*
+!analysis_out/georef_check_*/georef_check.json
+!analysis_out/georef_check_*/registration_*.png
+# The _chips variant re-renders the SAME two visual sites as its city's run, so
+# its PNGs are duplicates; only its JSON says anything new (the per-chip medians).
+analysis_out/georef_check_*_chips/registration_*.png
+
+# Ramps vs the city's own centrelines (§5i). Numbers only, no imagery — this is
+# the leg that exonerated Seattle's coordinates, so it has to be checkable.
+!analysis_out/centerline_offset_*.json
+
 # Claude Code local state. settings.local.json is per-machine (absolute paths,
 # session ids) and worktrees/ holds full repo checkouts -- gigabytes. Listed
 # individually rather than ignoring all of .claude/, so a shared
@@ -64,3 +114,15 @@ benchmark/*/index.csv
 # scripts/model_comparison/yolo_baseline/README.md, "Where the weights live").
 *.pt
 *.pth
+
+# Cached parquet columns for scripts/analysis/stage1_bearing_residual.py.
+# Derived from the published HF dataset on demand; regenerable with --refresh.
+analysis_out/stage1_bearing_cache/
+
+# At-the-sites basemap probes (#96): a few KB each, and they carry the paired
+# vegetation/coverage numbers a tile-source choice is justified by.
+!analysis_out/probe_sites_*.json
+
+# Inventory discovery sweep (#96): the candidate list, with record counts. A few
+# tens of KB, and it is the evidence that supply is not the constraint.
+!analysis_out/inventory_discovery.json

analysis_out/centerline_offset_denver-co.json

@@ -0,0 +1,114 @@
+{
+  "centerline_paths": 7966,
+  "centerlines": "denver-co-centerlines-2026-07-31.jsonl.gz",
+  "city": "denver-co",
+  "interpretation": "Measures the ramp coordinates against the SAME publisher's street geometry, so it is blind to any error the two layers share. A shift near zero means the coordinates are consistent with their own city's road network: any offset the reviewer saw against imagery is then in the frame (datum, reprojection or basemap registration) and is a constant that can be subtracted. A shift matching the reviewer's means the defect is in the ramp layer itself and no constant fixes it.",
+  "inventory": "denver-co-2026-07-31.jsonl.gz",
+  "limits": "A centreline is a cartographic construct, not a survey of the pavement midline, and this estimator assumes only that it is unbiased BETWEEN the two sides of the street -- it never uses its absolute position, which is what makes it robust to that. It cannot see an error the ramp and centreline layers share, by construction; that is what verify_chip_georeference.py is for. Cluster overlap attenuates the estimate toward zero when the half-width is small relative to the spread, so a near-zero reading is weaker evidence than a large one.",
+  "max_dev_deg": 20.0,
+  "max_dist_m": 25.0,
+  "near_cardinal": {
+    "east_west": 38547,
+    "north_south": 36081
+  },
+  "ramps": 72770,
+  "result": {
+    "axes": {
+      "east": {
+        "half_width_m": 7.348486962728202,
+        "iqr_neg_m": [
+          -8.911230517551303,
+          -5.747113466262817
+        ],
+        "iqr_pos_m": [
+          5.597882028669119,
+          8.774632066488266
+        ],
+        "median_neg_m": -7.466750402003527,
+        "median_pos_m": 7.230223523452878,
+        "n": 49594,
+        "n_neg": 24807,
+        "n_pos": 24787,
+        "shift_ci95_m": [
+          -0.13743286952376366,
+          -0.09783872496336699
+        ],
+        "shift_m": -0.11826343927532434
+      },
+      "north": {
+        "half_width_m": 7.459344879258424,
+        "iqr_neg_m": [
+          -9.21053040586412,
+          -5.859040271490812
+        ],
+        "iqr_pos_m": [
+          5.8847334291785955,
+          9.200584521517158
+        ],
+        "median_neg_m": -7.454281134530902,
+        "median_pos_m": 7.464408623985946,
+        "n": 49404,
+        "n_neg": 24644,
+        "n_pos": 24760,
+        "shift_ci95_m": [
+          -0.017781874863430858,
+          0.024741371627897024
+        ],
+        "shift_m": 0.005063744727522135
+      }
+    },
+    "bearing_deg": 272.451764426082,
+    "half_width_band_m": [
+      3.0,
+      20.0
+    ],
+    "half_width_plausible": true,
+    "resultant_m": 0.11837179807662734,
+    "shift_east_m": -0.11826343927532434,
+    "shift_north_m": 0.005063744727522135,
+    "sign_convention": "Positive east means the PUBLISHED RAMP POINT lies east of where the centreline geometry puts the street, i.e. the coordinate is east of the true ramp. This is the same sense the review sheet reports, so the two are directly comparable and must not be negated."
+  },
+  "segments": 114021,
+  "sweep": [
+    {
+      "max_dev_deg": 10.0,
+      "n_east": 47259,
+      "n_north": 46789,
+      "resultant_m": 0.12162942143892518,
+      "shift_east_m": -0.12154968967661262,
+      "shift_north_m": 0.004403305472806096
+    },
+    {
+      "max_dev_deg": 15.0,
+      "n_east": 48504,
+      "n_north": 48283,
+      "resultant_m": 0.12009124199139472,
+      "shift_east_m": -0.12005011225119233,
+      "shift_north_m": 0.003142761765047908
+    },
+    {
+      "max_dev_deg": 20.0,
+      "n_east": 49594,
+      "n_north": 49404,
+      "resultant_m": 0.11837179807662734,
+      "shift_east_m": -0.11826343927532434,
+      "shift_north_m": 0.005063744727522135
+    },
+    {
+      "max_dev_deg": 25.0,
+      "n_east": 51000,
+      "n_north": 50817,
+      "resultant_m": 0.11641682097028895,
+      "shift_east_m": -0.11626110132783651,
+      "shift_north_m": 0.0060193457175046206
+    },
+    {
+      "max_dev_deg": 30.0,
+      "n_east": 52417,
+      "n_north": 52392,
+      "resultant_m": 0.11152097366727215,
+      "shift_east_m": -0.11143287783488631,
+      "shift_north_m": 0.0044318512082099915
+    }
+  ]
+}

analysis_out/centerline_offset_seattle-wa.json

@@ -0,0 +1,114 @@
+{
+  "centerline_paths": 34484,
+  "centerlines": "seattle-wa-centerlines-2026-07-31.jsonl.gz",
+  "city": "seattle-wa",
+  "interpretation": "Measures the ramp coordinates against the SAME publisher's street geometry, so it is blind to any error the two layers share. A shift near zero means the coordinates are consistent with their own city's road network: any offset the reviewer saw against imagery is then in the frame (datum, reprojection or basemap registration) and is a constant that can be subtracted. A shift matching the reviewer's means the defect is in the ramp layer itself and no constant fixes it.",
+  "inventory": "seattle-wa-2026-07-31.jsonl.gz",
+  "limits": "A centreline is a cartographic construct, not a survey of the pavement midline, and this estimator assumes only that it is unbiased BETWEEN the two sides of the street -- it never uses its absolute position, which is what makes it robust to that. It cannot see an error the ramp and centreline layers share, by construction; that is what verify_chip_georeference.py is for. Cluster overlap attenuates the estimate toward zero when the half-width is small relative to the spread, so a near-zero reading is weaker evidence than a large one.",
+  "max_dev_deg": 20.0,
+  "max_dist_m": 25.0,
+  "near_cardinal": {
+    "east_west": 56178,
+    "north_south": 80645
+  },
+  "ramps": 38364,
+  "result": {
+    "axes": {
+      "east": {
+        "half_width_m": 6.241083346307278,
+        "iqr_neg_m": [
+          -8.479622146114707,
+          -4.798670271411538
+        ],
+        "iqr_pos_m": [
+          4.738980691879988,
+          8.389321817085147
+        ],
+        "median_neg_m": -6.241670722141862,
+        "median_pos_m": 6.240495970472693,
+        "n": 31430,
+        "n_neg": 15746,
+        "n_pos": 15684,
+        "shift_ci95_m": [
+          -0.0011285701766610146,
+          0.00015931949019432068
+        ],
+        "shift_m": -0.0005873758345842361
+      },
+      "north": {
+        "half_width_m": 6.256053883582354,
+        "iqr_neg_m": [
+          -7.782055872492492,
+          -4.2723733792081475
+        ],
+        "iqr_pos_m": [
+          4.224941612221301,
+          7.776318969205022
+        ],
+        "median_neg_m": -6.256016336381435,
+        "median_pos_m": 6.256091430783272,
+        "n": 30809,
+        "n_neg": 15463,
+        "n_pos": 15346,
+        "shift_ci95_m": [
+          4.033325240015984e-06,
+          7.086689583957195e-05
+        ],
+        "shift_m": 3.754720091819763e-05
+      }
+    },
+    "bearing_deg": 273.65757824963987,
+    "half_width_band_m": [
+      3.0,
+      20.0
+    ],
+    "half_width_plausible": true,
+    "resultant_m": 0.0005885746879966207,
+    "shift_east_m": -0.0005873758345842361,
+    "shift_north_m": 3.754720091819763e-05,
+    "sign_convention": "Positive east means the PUBLISHED RAMP POINT lies east of where the centreline geometry puts the street, i.e. the coordinate is east of the true ramp. This is the same sense the review sheet reports, so the two are directly comparable and must not be negated."
+  },
+  "segments": 243912,
+  "sweep": [
+    {
+      "max_dev_deg": 10.0,
+      "n_east": 29547,
+      "n_north": 29555,
+      "resultant_m": 0.00035881514778370217,
+      "shift_east_m": -0.0003569074906408787,
+      "shift_north_m": 3.695068880915642e-05
+    },
+    {
+      "max_dev_deg": 15.0,
+      "n_east": 30709,
+      "n_north": 30292,
+      "resultant_m": 0.0004877114404616942,
+      "shift_east_m": -0.0004863874055445194,
+      "shift_north_m": 3.591296263039112e-05
+    },
+    {
+      "max_dev_deg": 20.0,
+      "n_east": 31430,
+      "n_north": 30809,
+      "resultant_m": 0.0005885746879966207,
+      "shift_east_m": -0.0005873758345842361,
+      "shift_north_m": 3.754720091819763e-05
+    },
+    {
+      "max_dev_deg": 25.0,
+      "n_east": 32136,
+      "n_north": 31331,
+      "resultant_m": 0.0006898383046268158,
+      "shift_east_m": -0.000688822939991951,
+      "shift_north_m": 3.7414487451314926e-05
+    },
+    {
+      "max_dev_deg": 30.0,
+      "n_east": 33009,
+      "n_north": 31854,
+      "resultant_m": 0.0007770327259020268,
+      "shift_east_m": -0.0007764431647956371,
+      "shift_north_m": 3.0263327062129974e-05
+    }
+  ]
+}