test(svg): pin SVG disambiguation heuristics (fragile area #1) + latent bug

[jsboige]

What

Pin the SVG disambiguation heuristics (FallacyMindMapDocumentConfig.CollectPossibleSvgNodes + DisambiguateSvgNodes) — CLAUDE.md "Known Fragile Area #1" — with 7 contract tests via XElement fixtures. Additive only, no production code change. Full suite 429/0/5.

This is the ai-01 PRIMAIRE dispatch (msg-20260620T094434-yj9f75): "SVG disambiguation contract tests (fragile area #1)... Pin the behavior with XElement fixtures (pas d'extraction prod requise). Je confirme la valeur : risque silencieux #1 documenté."

Why

"SVG disambiguation in mind map generation" with "fragile heuristics dependent on Freeplane's output structure" is the #1 documented fragility (CLAUDE.md). A regression in the min-length grouping, the 3-char prefix fallback, or the parent-distance tie-break silently wires each fallacy to the WRONG mind-map node — caught only by visually inspecting every generated SVG.

These methods are PRIVATE instance methods bound to TitleFunc/Logger — audited and rejected for the #204 extraction lane (delegates + Logger + XElement coupling break the byte-identity-by-inspection bar). The coordinator's call: pin the BEHAVIOR with fixtures (no extraction), because the silent-wrong-output risk justifies the coverage regardless.

How — reflection + real Fallacy fixtures

The methods are invoked via reflection against real Fallacy instances (which implement IMindMapItem and expose TextFr, so TitleFunc's {item.TextFr} interpolation resolves cleanly). Minimal <g><text> XElement fixtures exercise each branch.

The 7 tests (SvgDisambiguationContractTests)

Test	What it pins
Collect_SingleExactMatch_ReturnsThatNode	happy path
Collect_MultipleMatches_DifferentLengths_KeepsShortest	THE fragile bit — min-length group wins (title node, not a re-quoting descendant)
Collect_NoExactMatch_LeavesItemAbsentFromResult	silent-drop contract (logged, never a null-value entry)
Collect_SubstringMatchAcrossTextElements_Resolves	string.Join over split <text> children (Freeplane label+detail)
Collect_TitleShorterThan3CharsAndNoMatch_ThrowsOnPrefixFallback	latent bug — guard-less Substring(0,3) on a <3-char title throws
Disambiguate_SingleCandidatePerItem_Passthrough	common path doesn't mutate selection
Disambiguate_EmptyInput_ReturnsEmpty_NotThrows	early-exit guard doesn't NRE on Descendants()

A latent bug documented (not fixed — output-neutral)

CollectPossibleSvgNodes line 1304 calls title.Substring(0, 3) unconditionally in the no-match branch. A title shorter than 3 chars with no exact match throws ArgumentOutOfRangeException. In practice fallacy titles are long enough, but the code has no guard — the test pins the throw so the hazard is documented (a future caller with a short label crashes loud here, not silently misbehaves). Fixing it would be a behavior change → separate PR.

Verification

• New tests: 7/7 pass.
• Full suite: 429 passed / 0 failed / 5 skipped — no regression (baseline 422 + 7 new).
• Build: 0 errors.
• No production code touched — pure additive coverage.

Related

• Dispatched by: ai-01 PRIMAIRE (msg-20260620T094434-yj9f75, "couverture tests — SVG disambiguation fragile area Working version of Arguments with sketch #1").
• Companion to the rejected-extraction audit (mindmap zone fails the test(pdf): #204 extract PdfManager filename-suffix insertion to pure method + 6 tests #553 byte-identity bar; behavior-pinning is the chosen path).
• DoD: additive tests only — mergeable without a content gate.

🤖 Generated with Claude Code