TCS → guardrails: recall-first integration plan

Goal: turn “about half are detected” into a frozen, re-runnable measurement and a taxonomy of why each rail is missed.

• Answered 2026-07-18 (Miro): the ≈50 % figure is Miro's visual observation, not a measured number — no ground truth or prior comparison/QC exists to trace it back to. There is nothing to "pin down"; the Phase-0 baseline must be built visually, from scratch. Starting denominator: the existing 367 w-beam-run audit on battlebox (2026-07-17 asphalt-edge research note) plus the reference-set overlays below.
• Reuse the already-designed blinded-window protocol from that note (252 random 20 m windows, 2/segment + 96 targeted, 10 % double-labeled, lateral buffers 0.25/0.5/1/2 m, locked 32-segment test set) instead of inventing tagging from scratch. Beware self-referential truth: same-cloud manual references were measured ~10 pp too forgiving on curbs; step7_xml/abschnitt_4_5 is empty and lane_state_snapshot.json is pipeline-derived — neither is independent.
• Render top-down + perspective overlays with current detections; tag every visible rail run (Opus for visual tagging, human spot-check on the double-labeled sample). Exclude historically flaky segments (Long-production failure list) from the reference set.
• Classify each miss into: (a) no returns / occlusion — the documented #1 published recall killer, (b) candidate starvation — points exist but never enter the candidate set, (c) DEM/HAG error — points enter but fall outside the 0.42–0.78 m rail band (p8 contamination under rail occlusion and slope terracing at break lines are both documented mechanisms), (d) killed by cluster gates (length / spacing / heading / qualifying ratio), (e) lost in merge/dedupe or occlusion bridging, (f) mislabeled as corridor exclusion, (g) record-boundary artifact gap — data artifact, excluded from the recall denominator.
• Deliverables: per-class counts, per-segment recall baseline, and a frozen exemplar list (segment/record/station ranges) checked into the repo for regression re-runs.

Gate: taxonomy exists; classes (b)+(c) quantified — these are the TCS-addressable classes the rest of the plan bets on. If (a) occlusion dominates, revisit targets before implementing.

Goal: measure the recall ceiling the residue offers before writing any detector code.

• Rail analogue of the wall-survival experiment: for every Phase-0 missed rail, compute the fraction of its neighborhood points present in ~kept_mask and inside the corridor (variants: elev_scalar 0.0 vs 1.25).
• Prototype (standalone script, no repo changes): candidates = current set ∪ residue-in-corridor → existing occupancy/DBSCAN parameters → count how many Phase-0 miss exemplars now form clusters, and how many new non-rail clusters appear (FP preview).
• Reuse the battlebox experiment scaffold from the wall run (/home/ai/tcs-wall-exp/ pattern; per-segment out dirs to avoid the run_summary clobber).
• Add the TCS embankment calibration grid already proposed in the 2026-07-17 note — window {3, 6, 12 m} × elev diff {0.05, 0.15, 0.30 m} × slope {0.15, 0.30} — on embankment-heavy segments: tablecloth defaults are lip-tuned for asphalt-edge and were never validated on >15 % grades, exactly where rails stand.
• Known residue blind spot to measure, not discover later: SMRF's ≤3 m window lets trucks/decks survive into the ground surface → suppressed HAG and missing residue near large vehicles; rails alongside parked trucks will not be rescued by the union.

Gate: residue union recovers a majority of class-(b)/(c) misses on the exemplars. Otherwise stop and re-plan around the dominant miss class.

Goal: ship the recall lever. Guardrails learns to consume TCS masks.

• New inputs: --tablecloth-masks-dir (per-record *_tablecloth_masks.npz); config flag default off, bit-identical when off.
• In detect._accumulate_candidates: candidate points = (HAG band ∩ corridor) ∪ (residue ∩ corridor ∩ sanity band 0.15–2.5 m). Sanity band uses the best available height reference (p8 now, SMRF surface after Phase 3).
• Recall lever in geometry.select_candidate_cells: for cells that are ≥~80 % residue-backed, widen the rail mean-HAG acceptance (e.g. 0.30–1.20 m instead of 0.42–0.78) — exact numbers tuned on the Phase-0 exemplars, not invented.
• TDD workflow (Red: Grok writes failing tests incl. bit-identical-off and vehicle-not-imported adversarial cases; Green: Sol implements; Review: Opus correctness / Sonnet constraints / Grok test adequacy / Sol second opinion; battlebox suite).

Gate: recall on the reference set rises by the Phase-1-predicted amount for classes (b)/(c); FP looks acceptable under visual QC (no numeric ceiling — no ground truth to measure one against; see Risks). Visual QC round on 5 segments.

Goal: fix miss class (c) at the root. Staged per the recorded 2026-07-17 decision (Option C preferred; native TCS only on escalation) — that decision was made for precision-era criteria, so the recall re-measure is the new deciding metric.

• 3a — Option C hybrid: build the existing 0.75 m p8 DEM from TCS-ground-only points (mask input, no grid change, all ~15 height-gate constants untouched, streaming contract preserved). Removes object contamination under rails/trucks/canopy. Adopt criteria: recall gain on class-(c) exemplars; churn on clean segments reported (the recorded <2 % criterion is re-evaluated for the recall goal, not blindly enforced).
• 3b — native surface (escalation only): if slope/embankment false negatives persist: tablecloth persists SmrfSurface per record behind --set surface_output=true (npz: grid, grad, xmin, ymin, cell_m, elev_scalar; version bump + Nexus); guardrails mosaics per-record surfaces onto the tile frame (cellwise min overlap), GroundModel adapter, config-gated with p8 fallback. De-quantizes the rail band (±25 cm raster error at 0.75 m vs 0.20 m). Memory note: 0.20 m segment-wide grids are ~10 GB-class at full-frame histogramming — keep the streaming path.
• Either stage: min(ground_model, p8) inside wall corridor bands; validation via label-free ΔDEM raster + HAG band-flip census across the 0.20/0.30/0.42/0.78/2.0 m gates on non-bridge segments (AI3D-339 audit pattern), plus recall re-measure on the reference set; segment-033 bridge-guard regression test mandatory.

Gate: class-(c) misses collapse on the reference set; band-flip census shows no mass migration of pavement/lip points; suite green on battlebox.

• Wall evidence = residue clusters (clean walls: bodies were 98.5–100 % captured) merged with the banded-cell channel (covers grazing-angle faces). Footing z from the surface sample, not residue bottom (footing bands measured 0–60 % removed — ambiguous by nature).
• Kept∩wall-candidate flag: kept-as-ground points inside strong wall-candidate cells → annotate instance for review. Diagnostics showed this flag catches bridge artifacts (segment-135 wall 2 = overpass deck top, a wall-channel false positive the residue correctly ignored) as well as potentially climbed terrain — route flagged instances toward the bridge-artifact rejection rule first, banded-cell recovery second.
• Disagreement QC into run_summary.json: removed∩HAG<0.2 m rate (lip erosion), kept∩candidate-cell rate (wall vanishing), removed-fraction outliers (vehicle-heavy records).

Gate: wall recall on 134/135 unchanged-or-better with no new FPs between edge lines (carriageway rejection rule still authoritative); QC fields present and sane across the reference set.

• Decide AML pipeline ordering: run tablecloth once upstream and pass masks/surfaces to both asphaltedge (cleaned) and guardrails (masks + surface) — one filter run, two consumers — vs on-demand inside guardrails. Default recommendation: upstream stage, cached like ACC LAS.
• Full non-bridge A4/5 sweep on battlebox; final recall/precision report vs the Phase-0 baseline; visual QC round.
• Wrap-up both repos (version bump, Nexus, Jira-tagged commits); PRs per phase merged in order.

Gate: headline number — per-length recall on the reference set — reported against the ≈50 % baseline, with FP judged acceptable via visual QC (no measurable ceiling).