fix(stage5): dedupe merged arrows by (src, op, dst) tuple

[RichardHightower]

Summary

Replace Stage 5's full-line-string arrow dedupe with a (src, op, dst) tuple key so labelled variants like A --> B : foo and A --> B : bar collapse to one canonical edge.

Why

Discovered during the agent-brain trust-eval pass (docs/gen/agent-brain/EVAL.md). The _merge_class_diagrams helper used a set[str] keyed on the full line, so two arrows with the same (source, op, target) but different labels survived as distinct entries. Live in docs/gen/agent-brain/packages/models/README.md:

BaseModel <|-- GraphQueryContext : extends
BaseModel <|-- GraphQueryContext : inherits

Loading

Five duplicate-edge pairs across the models overview alone (GraphQueryContext, GraphTriple, CodeChunkStrategy, QueryMode, JobRecord *-- JobProgress). The diagram still renders but the noise undermines the overview's job of being a clean bird's-eye view. This is the same class of issue I papered over in #60 with the post-strip case in config; the eval caught that it also fires without label stripping in models.

Closes #62.

Changes

• src/designdoc/stages/s5_mermaid.py
  • New _parse_arrow(line) — returns (src, op, dst, label) or None. Sorts _ARROW_OPS longest-first as defence against future ops that might become substrings of one another (no current op is, but it's free insurance).
  • _merge_class_diagrams replaces arrows: set[str] with arrow_keys: set[tuple[str, str, str]] + arrow_lines: list[str]. First label seen wins — deterministic given the caller's stable input order (sorted class_docs).
• tests/unit/test_stage5_package_diagrams.py
  • 7 new tests: parser correctness across all 10 relationship ops, label-variant collapse, distinct-op survival, distinct-target survival, malformed-line None return.
  • All existing tests (identical-arrow dedupe, arrow preservation, non-classDiagram skip, etc.) continue to pass without modification.

Invariants

• MAX_ATTEMPTS=3 preserved (loop.py untouched)
• Checker isolation preserved (no doer/checker context bridging)
• Schema-validated verdicts / fail-loud preserved
• Mermaid two-checker (mmdc + LLM) preserved — change is in post-processing only; both checkers still run against the per-class diagrams unchanged
• HIL fallback preserved

Verification

• task ci green locally (107 passed in 56.76s)
• 19 unit tests in test_stage5_package_diagrams.py pass (12 prior + 7 new)
• TWRC discipline followed (RED with ImportError → implementation → GREEN)
• Manual rerun against agent-brain target — will verify that packages/models/README.md loses the 5 duplicate-edge pairs (not part of merge gate; the unit tests prove the behaviour)

Related

• Closes Stage 5: dedupe merged arrows by (src, op, dst) — labelled variants survive #62
• Discovered via docs/gen/agent-brain/EVAL.md
• Related: Stage 5: label-strip retry for package-overview diagrams that fail mmdc #59 / feat(stage5): label-strip retry for package overview diagrams #60 (the post-strip duplicate case in config was a symptom of this same dedupe weakness; this PR fixes the root cause)