AI3D-339 — TCS ground filter + rule-based conifer detection (no ML training)

ml.py:9 calls qc_verdicts.json "the human QC verdicts". It is LLM output (Sonnet/Opus inspector batches + Fable adjudication, order-matched, detector @74d6942) — the docstring asserts a guarantee the data does not carry. Replace with an accurate provenance statement. Land first, standalone.

Done when: read back; no code path touched.

Dependency (resolved 2026-07-18): the tablecloth wheel is not on Nexus (verified: authenticated resolver reports not-found). Use path source via [tool.uv.sources] (commented precedent pyproject.toml:33-43); zero-setup on battlebox (sibling checkout + venv @ d612f67). Publish iolabs-point-cloud-tablecloth 0.1.0 to Nexus (wrap-up flow) before CI/Azure needs this branch — note it in the PR.

API: classify_ground(points, config) → kept_mask (tablecloth/ground.py:27), build_smrf_surface_streaming (:49), classify_against_surface (:100), classify_mask_from_chunks (:120); SmrfSurface = grid, xmin, ymin, cell_m (:17-24).

Mechanism: new module tcs_ground.py: per record, classify → write_cleaned_npz(source, dest, kept_mask) → filtered path list → repoint the DEM build only (detect.py:465-471).

critical Repoint the DEM input only — never the candidate input. accumulate_candidates (detect.py:491) keeps reading the original run3. TCS removes vegetation as non-ground: feed cleaned NPZ to the candidate path and every tree disappears, presenting as "TCS didn't help", not as a bug. Distinct names (run3 / run3_ground) + a call-site assertion.

Mechanism A/B: if the C-profile sweep leaves slope artifacts, csf_cloth has road-corridor-validated starting profiles per slope class (Cai 2018): flat → rigidity 3 / res 0.3 / thresh 0.15 / 150 iter; gentle → 2 / 0.2 / 0.15 / 200; steep (cuttings) → 1 / 0.3 / 0.2 / 300. Pick per segment terrain. CSF has a verified bridge-as-ground failure interacting with the bridge guard (trees.py:144-146, tuned on segment 033) — regression-check 033 whenever csf_cloth is tried.

Done when (all label-free, A4_5):

1. ΔDEM raster audit before any detector run — (Option-C DEM − baseline p8), histogram + map of |Δ| > 0.15 m cells; attribute each region: vehicle/canopy removal (intended) vs terracing/fill vs slope/wall climb (ground climbing vegetated slopes or retaining walls → HAG≈0 → objects vanish from residue; check cells adjacent to walls/noise barriers) vs crest cut-off (convex embankment tops cut from ground → false residue).
2. Band-flip census — recompute HAG for every current detection under the C DEM; count crossings of 0.30/0.60/0.70/0.90/1.5/4.5/6.0 m boundaries. Near-zero flips on clean segments + concentrated flips on contaminated ones = working as designed.
3. Per-class detection-count deltas on curated segments (volume, not correctness).
4. Flag off ⇒ DEM byte-identical.
5. Runtime/RSS benchmark; cleaned npz to local ext4, not /mnt/d.

Pure declarations, zero logic — every multi-agent file, landed once, alone.

• config.py: TCS block (enable flag + profile params); conic-gate block (enable flag default off + all thresholds as config fields). Each field needs its loader line (tree_detection.get(…) pattern, config.py:617-646).
• features.py:15-59: declare new ClusterFeatures fields, 0.0 defaults, no computation: taper_slope_n, taper_ratio, apex_deg, h_over_width, azimuthal_coverage, crown_h_over_w, decile_fill_fraction. No verticality_raw, no greenness_exg. Names are the contract.
• detect.py:152-190 (_features_to_row): add every new field to the dict literal — otherwise the features never reach clusters.csv and the distribution dumps are empty.
• NOT touched: ml.py, TREE_FEATURE_NAMES, existing greenness/verticality semantics.

Done when: full pytest green; one real A4_5 segment shows new columns as 0.0 in clusters.csv; the shipped veg bundle loads and scores byte-identically.

The synthetic table IS the fixture — tests first, from idealized cone/cylinder/sphere samplers. Assert direction and ordering, not margins (samplers have no occlusion/clutter; real margins will be weaker):

feature	cone	pole	broadleaf
taper_slope_n	≈ −1.03	≈ 0.00	≈ +0.96
apex_deg	10.6–29.0	0.8–1.9	38.7
h_over_width	1.9–5.4	30–70	1.4

Method (_conicity(xy, heights, h_min, h_max)): zn ∈ [0,1], 10 deciles; per-decile radius = 90th pct of XY distance from the decile's median XY (median → robust to one-sided scan; percentile-of-distance → robust to far-side occlusion). Guard: zeros if <5 deciles have ≥10 points. Vectorize via bincount per _core_rms_median (features.py:77-113). Emit decile_fill_fraction alongside. azimuthal_coverage: implement (16 sectors above 0.5·H), expect nothing.

Not doing: any change to _verticality, greenness, or any existing field (hard additive constraint); any grid.py change.

Done when: TDD table green; regression tests pin verticality/greenness byte-identical on synthetic clusters; distribution dump of the new columns across A4_5 tree-path clusters — including change_of_curvature/omnivariance for the texture-conjunct calibration (do real values bimodally separate around the seed thresholds, or smear? feeds the Phase-5 enable decision).

§1a — dedup type filter ~3 lines: dedup_trees_against_signs (trees.py:180-198) currently drops any tree within 2.0 m of any sign-path detection — including the pole_other the sign path itself emitted for that conifer (classify.py:195). Filter suppressors before sign_xy is built (trees.py:188-190):

SUPPRESSING = {"sign", "delineator", "gantry_or_gate"}   # NOT pole_other / sign_post

Type strings per detect.py:218-222; sign_post is a distinct internal type (reported as pole_other in the sidecar, detect.py:227-232) — exclude both. Don't touch the radius or detect.py:718-720.

Conic OR-bypass — at the emit decision (trees.py:147-150): emit if confidence ≥ threshold OR (gate enabled AND conic rule passes). Scope change vs rev 4: the rule is two-cue by default — shape (taper) AND surface texture — because the verified literature refutes single-cue-family vegetation rules (foliage can read linear or planar at a single scale; every validated rule system requires ≥2 independent cues). Foliage is scattering-rough; a pole/mast is smooth. All thresholds from config; seed values:

shape   = (f.taper_slope_n <= -0.4 and 5 <= f.apex_deg <= 35
           and 1.5 <= f.h_over_width <= 12)
texture = (f.change_of_curvature >= cc_min or f.omnivariance >= omni_min)
           # existing fields — rules may read them today; calibrate both
           # thresholds from Phase 3a's A4_5 distribution dump
conic = (shape and texture
         and f.hi_intensity_fraction < 0.2
         and f.h_max >= 2.5 and f.on_road_fraction < 0.6
         and f.decile_fill_fraction >= fill_min)   # abstain on occlusion-starved profiles

seed values Unvalidated; must not ship enabled by default. Upstream hard gates still apply (tree_min_points, tree_crown_h_min_m, tree_max_dist_to_road_m, bridge guard trees.py:144-146) — the bypass skips only the RF. Emissions get reason="conic_rule" + experimental: True. A texture_cue_enabled config field (default on) allows A/B-ing the texture conjunct off during Phase 5 diagnostics. If A4_5 distributions show change_of_curvature/omnivariance don't separate foliage from poles at cluster level, the fallback texture cue is a per-cluster sphere/cylinder neighbor ratio (planar/solid → ~100, canopy low; pure XYZ, no return numbers needed) — that variant needs a new feature field, so it goes through a contract-commit addendum, not ad hoc.

Leave alone (verified not miss causes): seed gate (trees.py:33-50), tree_crown_max_area_m2=4.0, tree_dedup_radius_m=2.0, tree_eps_m=1.5, tree_crown_h_min_m=2.0.

Done when: unit tests — dedup: tree at 1.0 m from pole_other survives, from sign dropped, empty list short-circuits; bypass: conic synthetic emits at sub-threshold confidence, pole synthetic does not, occlusion-starved cone abstains, gate-off ⇒ identical to today. Then A4_5 curated segments: tree-count delta from §1a alone, then §1a + gate, counts split by reason.

Hard barrier; single agent owns the fix pass. Full pytest on battlebox.

Acceptance gate for additive-only: all new flags off ⇒ end-to-end output on 2–3 real A4_5 segments byte-identical to baseline except the §1a dedup delta. Any other diff is a bug — especially in greenness/verticality columns or sign-path detections.

• ΔDEM audit + band-flip census (Phase 1 items 1-2).
• Detection-count deltas per class × curated segment × {baseline, +§1a, +gate, +TCS}.
• Overlay inspection — agent-first. Generate --dump-point-masks + per-cluster overlays for every conic_rule emission and every band-flipped detection on the 13 curated segments. The agent views these itself and judges conifer-vs-pole and gate-flip sanity: read the overlay PNGs directly for the quick pass; for anything ambiguous, open the cluster's actual point cloud in CloudCompare on battlebox (Codex computer-use via codex-bb-runtime, optional — desktop stays on battlebox, local screen free), rotate/compare, and capture screenshots or a short video to keep as the record. This agent pass is the acceptance bar. Any fault the agent sees — a pole tagged conic_rule, a mis-flipped band, a bad threshold — it fixes immediately (tighten the gate, correct the code) and re-runs, no human second opinion required. The saved overlays/screenshots double as a browsable review view for optional human escalation after the agent's pass.
• Conic-gate distribution check (Phase 3a dump) — if real A4_5 values smear across the seed thresholds instead of separating, the gate ships off, period.

Then adversarial review (Opus — shipped-RF OOD surface + sign-path regression; Sol — _conicity numerics; Grok — the run3/run3_ground seam + bypass logic), serialized fix pass, re-run the Phase-4 byte-identical check, draft PR on the working branch.