Evaluate relaxing node_straddles containment now that on_minor_arc is SoS-hardened

[espg]

🤖 from Claude

Tracking follow-up — awaiting triage. No action label applied yet; leaving label/triage to the maintainer.

Context

PR #87 (Refs #78) made robust_crossing's antipodal-disambiguation step on_minor_arc exactly SoS-hardened: each wedge sign is now decided on a canonical, identity-sorted evaluation with an identity-keyed symbolic perturbation tie-break, so the predicate is total and reorder-invariant with no residual f64 sign sensitivity at endpoint coincidence.

Before that change, the caveat at src_rust/src/coverage.rs:442 noted that boundary cells affected by on_minor_arc's residual float sensitivity were caught by the node_straddles over-refinement safety net — such cells are refined (descended into) rather than filled whole. Now that the disambiguation is exact, that safety net may be loosenable: some cells the descent currently over-refines for safety might be safely fillable, reducing output size / work.

Re-scope (post PR #106 / issue #103) — supersedes the plan below where they differ

PR #106 replaced the entire crossing pipeline with the uniform symbolic predicate (arcs_cross_sos + exact determinant signs), so the node_straddles over-refinement is no longer load-bearing for correctness anywhere — it is now pure conservatism, and this issue's measure-then-relax program can be run safely. Adjustments:

1. Baseline = PR Uniform symbolic crossing predicate: fix base-cell-boundary mis-fill #106's branch (or main once it merges), not pre-SoS-harden on_minor_arc in robust_crossing #87 main. The measurement is only meaningful against the exact predicate.
2. Relaxation target = the non-touching straddle over-refinement only. The closed-set contract from morton_coverage / from_wkt mis-fill polygons whose edges lie on base-cell boundaries (lon ≡ 0 mod 45°, lat = 0) #103 deliberately refines and includes boundary-touching cells (edge_hits_cell_edge's exact-incidence path); those refinements are contract, not waste, and must be excluded from the "over-refinement" count.
3. Guard rail: PR Uniform symbolic crossing predicate: fix base-cell-boundary mis-fill #106's parity oracle (debug_assert in the descent) validates every uniform cell against the independent winding PIP — any relaxation experiment runs under it in the test profile, which the original plan lacked.

Recommendation

Measure before changing behavior. Specifically:

1. Quantify the over-refinement node_straddles currently causes (count of cells refined-not-filled, excluding closed-set boundary-touch refinements, across representative polygons) against the exact-predicate baseline.
2. Only then decide whether/how to relax the containment, with the measurement as the baseline to confirm no coverage correctness regression.

Do not change descent/containment behavior blind — the current over-refinement is conservative-correct, so any relaxation must be justified by a measured win and guarded by tests (and the parity oracle).

Refs #87
Refs #78
Refs #103, PR #106

[espg]

🤖 from Claude

Re-scoped the body per the #106 discussion (original text kept; a "Re-scope" section now supersedes it where they differ). Summary of what changed and why:

• Unblocked, not obsoleted: PR Uniform symbolic crossing predicate: fix base-cell-boundary mis-fill #106's exact predicate removes the correctness dependency on the node_straddles safety net, so measuring and relaxing it is now safe — before Uniform symbolic crossing predicate: fix base-cell-boundary mis-fill #106 the net was papering over the crossing predicate's residual float sensitivity and relaxation risked exposing wrong fills.
• Baseline for the measurement moves to the Uniform symbolic crossing predicate: fix base-cell-boundary mis-fill #106 branch (or main after merge).
• Scope of "over-refinement" now excludes the closed-set boundary-touch refinements added deliberately by morton_coverage / from_wkt mis-fill polygons whose edges lie on base-cell boundaries (lon ≡ 0 mod 45°, lat = 0) #103's contract — the relaxation target is non-touching straddle conservatism only.
• Guard rail: run any relaxation experiment under Uniform symbolic crossing predicate: fix base-cell-boundary mis-fill #106's parity oracle (active in the test profile).

Still unlabeled — triage remains yours; it's ready for a plan or implement label whenever you want the measurement run.

---

Generated by Claude Code

[espg]

🤖 from Claude

Implementation plan (measure → decide → relax), per the re-scoped body

Phase 1 — cause-tagged instrumentation (feature-gated, zero release cost)

node_straddles currently answers only whether a cell straddles; the measurement needs why. Add a cause taxonomy behind a descent-stats cargo feature (#[cfg(feature = "descent-stats")], off by default — no runtime cost in normal builds, nothing new on PyPI):

• VertexLeaf — a polygon vertex's leaf falls in the cell (clause 1);
• QuadCross — a relevant edge crosses a cell edge (the genuine boundary crossings);
• QuadTouch — the closed-set exact-incidence branch (edge_touches_cell_edge_degenerate) — excluded from the over-refinement count by contract (morton_coverage / from_wkt mis-fill polygons whose edges lie on base-cell boundaries (lon ≡ 0 mod 45°, lat = 0) #103);
• CornerParity — the centre→corner clipped-corner clause;
• NearPoleBulge — the densified morton_coverage: polygon-edge interpretation differs from S2 / shapely geodesic #32 path.

Counters aggregated per descent (thread-local, merged), exposed through one test-only binding. Files: src_rust/src/coverage.rs (tagging), Cargo.toml (feature — a cfg flag, not a dependency), one dev script under benchmarks/ (not wired to CI).

Phase 2 — the measurement itself

For each straddle-stopped leaf (refined to target order, emitted as boundary), independently test whether the polygon boundary actually intersects the cell: dense great-circle sampling of every relevant edge against the cell's true (densified) boundary, plus the exact predicate on the samples. A straddle leaf the boundary provably misses = true over-refinement; ambiguous cells are counted separately, not claimed.

Polygon suite × orders {6, 9, 11}: the #103 matrix shapes (on-grid boxes — the closed-set touches should dominate and be excluded correctly), the bench polygons (circles 32/100/500, mid-lat + polar triangles/squares), a synthetic ICESat-2-style swath, and one hemisphere+ ring. Report per (polygon class, order, cause): straddle-leaf count, provably-over-refined count, and the achievable output-size reduction (flat and MOC) if those cells were filled/absorbed instead.

Deliverable: a results table as a comment here, with the raw script committed to the measurement branch only ("run and reported, not merged" — the #78 benchmark precedent).

Phase 3 — decision + relaxation (separate PR, only if justified)

With numbers in hand, relax by cause, narrowest first (likeliest candidate: CornerParity cells whose quad test found no crossing and whose centre fill is true and boundary clearance exceeds the circumradius — fill whole instead of refining). Every candidate runs under the parity oracle (test profile), the full #103 matrix, plus a new no-under-coverage sweep comparing against dense-sampled ground truth. If phase 2 shows the win is small (see open question 3), stop at the report and close.

Acceptance criteria

• A quantified table on this issue: over-refinement by cause and polygon class, with the closed-set touches separated out.
• An explicit go/no-go recommendation with expected cell-count and wall-time deltas.
• If "go": the relaxation lands as its own PR with the guard rails above; this issue closes on the report either way.

Open questions (one pass)

1. Instrumentation form — cargo feature (descent-stats, my recommendation: usable from Python dev builds) vs #[cfg(test)]-only (Rust-side measurement only)? No new dependency either way.
2. KPI — output cell count (flat + MOC size), wall time, or both? I'd optimize for MOC size (zagg's storage cost) and report time as secondary.
3. Go threshold — what saving justifies the relaxation's risk surface? Proposal: proceed only if ≥10% MOC-size reduction on at least one realistic polygon class; otherwise document-and-close.

Applying waiting for your read on the three questions (or a straight go-ahead — phase 1 has no design forks).

---

Generated by Claude Code

[espg]

Instrumentation form

cargo feature (descent-stats, your recommendation: usable from Python dev builds)

KPI — output cell count (flat + MOC size), wall time, or both? I'd optimize for MOC size (zagg's storage cost) and report time as secondary.

both, with MOC size the priority

Go threshold — what saving justifies the relaxation's risk surface? Proposal: proceed only if ≥10% MOC-size reduction on at least one realistic polygon class; otherwise document-and-close.

yes, ≥10% MOC-size reduction or comparable wall time percentage reduction