Heuristic correctness: canonical fixtures and independent oracles

Parent: #17

Depends on the evaluation scaffold issue.

Outcome

Validate the mathematical and resolver correctness of matlatl's shipped mechanisms without using matlatl output as its own oracle.

Scope

• Canonical empty, singleton, path, star, cycle, SCC, DAG, disconnected, multi-root, bridge, articulation, and symmetric-tie graph fixtures.
• Resolver fixtures for relative, root-absolute, directory, fragment, alias/wikilink, ambiguous, external, non-note, stale, malformed, and root-escape references.
• Hand-enumerated or independently calculated expectations for reachability, bow-tie classes, hops, HITS, PageRank, navigability scalars, betweenness, articulation points, bridges, trails, backlinks, scent, knowledge gaps, and suggested links.
• Seeded reversible graph and reference mutations with expected deltas.
• Byte-stability and shuffled-input checks.

Constraints

• Binding definitions are the ADRs listed in docs/research/heuristic-evaluation.md.
• Any unexplained correctness mismatch blocks usefulness or agent-outcome claims for that mechanism.
• Floating-point comparison and accumulation order must be explicitly pinned.

Acceptance

• Every shipped mechanism has at least one independent correctness oracle.
• Threshold boundaries, exemptions, caps, tie ordering, and indeterminate cases are covered.
• Invalid universal monotonic assumptions are not encoded as tests.
• Repeated and shuffled-input runs are byte-stable.
• The complete offline correctness suite runs through task eval:oracle.

Start with the binding definitions in docs/research/heuristic-evaluation.md and review the evaluation scaffold dependency from issue #17. Build the canonical graph and resolver fixtures with independent expectations, then run task eval:oracle. Done means every shipped mechanism has an oracle, boundary and indeterminate cases are covered, and repeated or shuffled-input runs are byte-stable.