Hierarchy-mapping heuristics (dump path → design path): each tie-break 'fix' leaked a new wrong-module case until a randomized invariant sweep was added

Context: a waveform tool maps a dumped signal path (tb.dut.u_core1.u_a.x) to the elaborated design path (soc.u_core1.u_a.x). It infers rename rules (dump prefix ↔ design root) by scoring how many child names overlap. Candidate roots are the tops plus deeper instances.

What happened over 4 review rounds: 1. Adding deeper instance roots let two identical module copies tie, and first-by-name won: the row mapped to the wrong core. 2. Treating any tie as ambiguous broke the case where the tied rules map to the same path (a root and its own child instance), giving 'not found'. 3. Grouping ties by resulting path, then dropping to the next score, picked a lone, narrower rule at the same prefix. A different module that shares child names (u_dma with u_a/u_b) got chosen. That signal exists, so the trace ran silently on the wrong RTL. 4. Final rule: after an ambiguous score, accept a lower-scoring rule only if its dump prefix is strictly shorter than every tied rule's prefix. A rule at the same or a deeper scope is just another guess for the unresolved scope.

What finally settled it was not another example test but a randomized sweep. Generate 3-7 modules with child names drawn from a small shared pool, so names collide and module copies are common. Dump the same hierarchy under renamed prefixes or wrapper levels, resolve every leaf, and assert 'own path or no answer, never another instance'. Over 135k rows: final rule 0 wrong; the previous attempt 203 per 55k; first-wins ~1,600-2,500 per 40k.

Lessons: - For scoring/tie-break heuristics, state the invariant (never map to a different instance) and fuzz it, before fixing individual counterexamples. - Keep the old behaviour as an explicit mode for callers where any match is equivalent (e.g. a source-link lookup where all instances share one declaration). - Measure coverage ('no answer' rate) alongside wrongness. A safer rule can leave more rows unresolved. - A dense name pool overstates absolute rates; compare variants against each other.