Files

Tarik Moussa f25174ed69 feat: output_uv_map for InversiveDistance, error for CP-Euclidean, reviewer trio

Three reviewer-meeting deliverables in one commit.

(1) output_uv_map for the two remaining DCE entries
    ─────────────────────────────────────────────────
    * Discrete_inversive_distance.h: full implementation.  After Newton,
      reconstruct effective Euclidean edge lengths from the converged
      log-radii via the Bowers-Stephenson identity
      `ℓᵢⱼ² = rᵢ² + rⱼ² + 2·Iᵢⱼ·rᵢ·rⱼ`, populate a temporary
      EuclideanMaps with `lambda0 = log(ℓᵢⱼ²)`, and reuse the existing
      `euclidean_layout(mesh, 0, eucl)` priority-BFS.  Per-vertex
      Point_2 coordinates written into the user-supplied pmap.
      Optional `normalise_layout(true)` applies the canonical PCA
      centroid + major-axis rotation, same as the other 3 entries.

    * Discrete_circle_packing.h: throws std::runtime_error with a
      clear pointer to Phase 9c rather than silently producing
      nonsense.  CP-Euclidean is face-based; the faithful output is a
      per-face circle packing in ℝ², not a per-vertex Point_2 map.
      A true layout requires BPS-2010 §6 (~150 lines, on the porting
      roadmap as Phase 9c).  Failing loudly is the honest default.

    Tests: 2 new cases in test_cgal_phase8b_lite.cpp
    (OutputUvMap_InversiveDistance_PopulatesPmap;
    OutputUvMap_CPEuclidean_ThrowsClearly).  Both green.
    Suite total now 259 (was 257, +2).  CGAL subtotal: 234 → 236.

(2) Reviewer meeting documents
    ──────────────────────────
    New directory doc/reviewer/ with three files:

    * briefing.md  — one-page orientation for the reviewer.
      What the project is, where to look first
      (https://tmoussa.codeberg.page/ConformalLabpp/), the headline
      evidence (tests/coverage/sanitizers/license), what we want from
      them, what's deferred and why, and the 5 questions in a separate
      file.

    * questions.md — the 5 concrete decisions we want their second
      opinion on:
        Q1 Phase 9c (port-literal vs re-derive)
        Q2 Phase 9b-analytic (worth ~2 weeks for ~6× speedup?)
        Q3 CP-Euclidean output_uv_map (build now or defer?)
        Q4 CGAL submission strategy (one package or five?)
        Q5 geometry-central cross-validation co-authorship
      Plus an explicit "what would you say no to?" question at the
      bottom — negative feedback is the highest-value information.

    * agenda.md   — my own internal playbook (NOT to be sent).
      60-min flow: 5-min thank-you, 10-min architecture tour,
      30-min for Q1-Q5 in the order Q4-Q1-Q2-Q5-Q3, 5-min "no"
      question, 5-min wrap-up.  Includes post-meeting memo template
      to fill out in the 30 min after.

    * README.md    — index for the directory; says which file goes
      to whom and when to send.

(3) locked-vs-flexible.md known-limitations update
    ─────────────────────────────────────────────
    "output_uv_map covers 3 of 5 entries" → "covers 4 of 5".
    CP-Euclidean's throws-clearly behaviour documented as a Phase 9c
    deliverable rather than a passive gap.

Bonus: extended .codespellrc ignore list (acknowledgement, the
British-English spelling I used in agenda.md).

Verifications on this commit:
  259/259 tests pass (0 skipped)
  scripts/check-test-counts.sh:                OK (23 + 236 = 259)
  scripts/quality/license-headers.sh:          OK (66/66 SPDX)
  python3 scripts/quality/cgal-conventions.py: OK (0/6 violations)
  scripts/quality/codespell.sh:                OK (0 typos)
  scripts/quality/shellcheck.sh:               OK (0 findings)
  python3 scripts/check-markdown-links.py:     OK (143/143)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-05-26 11:15:09 +02:00

5.5 KiB

Raw Blame History

Questions for the external reviewer

Please skim this before the meeting. Each question is scoped so that "do A" / "do B" / "either is fine" is a sufficient answer; deeper dives are welcome but not required.

These are the five decisions that would benefit most from a second opinion. Items 1-3 affect the porting roadmap directly; 4-5 affect the long-term goal (CGAL submission).

Q1 — Phase 9c (4g-polygon canonical form): port-literal vs re-derive?

The Java original has two utility classes:

FundamentalPolygonUtility (~600 lines)
CanonicalFormUtility (~900 lines)

that together compute the canonical 4g-polygon for a higher-genus surface from its cut graph.

We can either:

(A) Port literally (~2 weeks). Faithful, predictable, fixes the Java algorithm in C++. Down-side: we inherit the Java code's ad-hoc style and edge-case handling.
(B) Re-derive from Springborn 2020 §5 (~3 weeks). Uses our existing cut_graph.hpp + holonomy infrastructure cleanly. Down-side: longer; potential for new bugs not seen by the Java original's test cases.

Question: which route do you prefer, and is there a Springborn-era reference implementation (Mathematica notebook, paper appendix) we should cross-validate against?

Context: doc/roadmap/phases.md §Phase 9c, doc/roadmap/porting-status.md.

Q2 — Phase 9b-analytic Hessian: implement now or later?

We have:

Done: per-face block-FD Hessian (96× faster than naive full-FD).
Derived but not implemented: analytic Hessian via the Schläfli identity, expected ~6× further speedup over block-FD.
Derivation document: doc/math/hyperideal-hessian-derivation.md (805 lines, all sign pitfalls covered, references Schläfli 1858, Milnor 1982, Vinberg 1993, Cho-Kim 1999, Glickenstein 2011, Springborn 2020).

Question:

At what mesh size does the ~6× become user-visible enough to justify ~2 weeks of implementation work?
Are you aware of subtleties in the Schläfli-based derivation our document might be missing?

If the answer is "implement", we'd target Phase 9b-analytic right after the meeting.

Q3 — `output_uv_map` for CP-Euclidean: build now or defer?

For CP-Euclidean (face-based DOFs, BPS 2010) the natural output is a per-face circle packing in ℝ² — not a per-vertex Point_2 map the way the other four DCE entries produce.

Today the entry throws std::runtime_error with a helpful pointer if the caller passes output_uv_map(...). A faithful layout would be ~150 lines implementing BPS-2010 §6.

Question: do your CP-Euclidean use cases need a UV-like output, or are the face circle radii themselves the deliverable? The answer shapes whether Phase 9c gets it now or later.

Q4 — CGAL submission strategy: one package or five?

For the long-term CGAL submission:

(A) One package "Discrete_conformal_map" — single entry header, five solver functions, one set of named parameters. Easier for users to find; harder to compartmentalise reviews.
(B) Five packages "Discrete_*" — each DCE model is its own CGAL package with its own concept + reference manual. More ceremony for users; more familiar review surface for CGAL editors.

Today the code is structured per-functional (Strategy C — see locked-vs-flexible.md §7). Either submission packaging is achievable from this base.

Question: what's the CGAL editor convention for related-but-distinct algorithms — Polygon_mesh_processing as one example (one package, many algorithms), vs Triangulation_2/Triangulation_3/Periodic_/Hyperbolic_ as another (multiple packages for related algorithms)?

Q5 — geometry-central cross-validation (GC-1)

Two libraries solve the DCE problem from opposite algorithmic directions:

conformallab++: Newton on the fixed mesh (no intrinsic flips).
geometry-central: Ptolemy flips on an intrinsic triangulation.

An automated comparison on common meshes would be:

A nice paper (the disagreement modes are interesting in their own right).
A confidence-building tool for both libraries (we benefit from catching corner-case bugs, they benefit from cross-validation against a Newton baseline).
~3 days of plumbing (CMake-fetch geometry-central, write 5 common test meshes, compare u_per_vertex to a tolerance).

Question: would you be interested in co-authoring such a comparison note (with us doing the implementation)? Or do you know someone in the Crane group who would?

Things you do not need to comment on (unless you want)

C++ style choices captured in .clang-format + .clang-tidy.
Test framework choice (GTest).
License (MIT, with vendored deps catalogued in code/deps/THIRD-PARTY-LICENSES.md).
Build system (CMake ≥ 3.20, header-only consumer + optional CLI/Viewer).

These are conscious decisions matched to CGAL conventions and aren't load-bearing in the sense that revisiting them later is cheap.

What I'm hoping you'll say "no" to

This is a deliberately blunt question because positive feedback is nice but negative feedback is rarer and more valuable:

Looking at any of the 12 architecture decisions in locked-vs-flexible.md, is there one where you think "no, that's the wrong call, here's why"?

The 🔴 load-bearing decisions are the most consequential to revisit, because waiting longer makes them more expensive. No is the most useful answer.

5.5 KiB Raw Blame History Unescape Escape