E01 · F02 · AUD-002 — Audit F02: regular polygon with corner radius

Full check of F02 in depth and of the project in breadth, following audit.md (ADR-0020), with the /audit skill.

Acceptance criteria

• Given the functions of F02, when audited, then every formula is re-derived, every expected value recomputed by hand, and each function survives a mutation spot-check.
• Given the sources, when audited, then every @see of F02 is re-verified online and dated.
• Given the project, when audited, then no duplicate, no copied code and no inconsistency between code, domain, ADRs and docs remains unrecorded.

Tasks

One task = one commit, referenced as AUD-002.Tn.

• T1 — A. Independent review by the auditor subagent; mathematics, mutation spot-checks, missing properties (3 h)
• T2 — B. Sources re-verified (1 h)
• T3 — C, D, E. Provenance, duplicates, consistency (2 h)
• T4 — F, G. Quality gates and research list for the next feature; findings table (1 h)

Findings

Independent review by the auditor (A–E), mutations and breadth checks by the agent, 2026-10-09.

Check	Result	Evidence	Action
A1 — derivation re-derived without the code	correct	steps 1–6 and the five check rows recomputed (node)	—
A1 — "within EPSILON" in step 4 and fitInBox	wrong for large boxes	error about 1e-16 × the size: 3.7e-9 at 1e8, −0.5 written at the largest safe integer	wording fixed (relative error); size range asked: Q20
A2 — expected test values	correct	every value recomputed by hand or with node	—
A3 — missing cases	pentagon, n = 12, rounded size 0	no example test	tests added
A3 — signed zeros	none	unitRegularPolygon, fitInBox from size 0 to the largest safe integer	—
A4 — mutations (agent)	10 of 10 caught	one per function: start index, minX, left, >= 0, integer check, max: 13, width/height swap, radius 0, extra cap, radius dropped	—
A4 — mutations (auditor)	16 of 17 caught	effectiveCornerRadius min → max survives	equivalent: the new property proves every corner has the same effective radius
A5 — properties	present	all geometry and domain functions of F02; predicates by examples	property "same effective radius" added
B1 — sources re-opened	all read again	MathWorld ×4, OpenStax, Euclid I.8 and I.47, Inkscape, ISA InTech	IEC, ANSI, ISO refuse robots (403): stated, link check excludes them
B2 — link check	green	lychee on the branch: 336 links, 0 error, 4 excluded	.lycheeignore rows with reason
B3 — @see to [unverified]	none	lint; no eslint-disable added	—
C — provenance	own	TSDoc formulas match bodies; index-based start and 0 − sin guard are project-specific	—
D — duplicates	one left	expectPoints in two test files	improvement log (shared fixtures)
E1 — playground help	wrong for polygons	"the curve starts that many pixels before the corner": true on a rectangle only	text fixed
E1 — research note	contradiction	pre-research line said IEC was not read	marked superseded
E4 — feature plan, testing convention	incomplete	"Verified by" and the playground test's criteria	completed
F — quality gates	green	check:all, coverage 100 %, npm run build, no .skip / .only	—

G. Next feature (F03, per-vertex radius and node editing)

Sources and questions for its spike:

• Per-vertex radius: REF-FIGMA-CR (verified) for the behavior; the clamp of F01 already works per vertex (Corner); the Math.max/Math.min mutant of effectiveCornerRadius stops being equivalent — the interface needs one effective radius per vertex.
• Node editing (select, move by integers, add on an edge, delete): Inkscape's node tool and Figma's vector editing as functional references, to read (no row yet); projecting a click onto an edge (closest point on a segment) needs a source.
• Concave contours: the known limit of the clamp (an arc touching a non-adjacent edge, shapes.md §2) becomes reachable — segment–arc and arc–arc intersection (REF-SCHNEIDER-EBERLY, REF-BOURKE-CIRCLES, both to verify).
• Business questions: can a regular polygon's vertices be edited (it would stop being regular: convert it to a free contour?); does a moved vertex keep its radius; what does deleting a vertex down to two do.
• Large boxes: the absolute EPSILON of hasLength leaves tiny segments on boxes of millions of units (improvement log) — a size range or a relative tolerance, for the Product Owner.

H. Light evolvability check (2026-10-09)

Question	Answer
Size	53 modules in src/ (geometry 21, math 11, model 11, io 6, render 4); 306 tests in 4.8 s; check:all about 28 s; no layer close to a package split
Tools	StrykerJS still blocked (issue 6210 open): mutations by hand, 10 for 10 caught in this audit; knip waits for the retrospective; deps:tools added at the Product Owner's request
Fitness functions	every new geometry function has a property test; still a candidate automatic check (improvement log)
Patterns	duplicated size rule removed (isModelSize); test helpers (expectPoints, signed area) logged; regularPolygonCorners mirrors rectangleCorners — fine for two shapes, a shared "corners of a shape" when a third comes
Agent configuration	CLAUDE.md + rules 185 lines; one library rule added (visibility in the playground); /run orders the review before done
Process	6 stories merged by the run; every review changed its story; one real bug caught only by review (hidden field), which happy-dom could not see before the head was loaded — real-browser tests remain an E12 item