Add canonical graph6 decoding and verification

[yuelgrace1810-ops]

Summary

• add graph.encoding.graph6.decode.compute for standard small-order graph6 (orders 0–62), with optional >>graph6<< header
• implement the standard column-major upper-triangle bit order and return canonical sorted edges, degree sequence, and a digest of the exact decoded graph
• reject sparse6, digraph6, extended-order headers, invalid lengths, invalid characters, and nonzero padding
• add graph.encoding.graph6.decode.verify with independent standard-library bitstream replay

Reproduction

The H24 string from #947 decodes to 24 vertices, 30 edges, and maximum degree 3. The regression fixture checks the first canonical edges and independent verifier, and a wrong-edge candidate is rejected.

Scope boundary

This PR implements only the deterministic codec slice of #947. It does not add or broaden maximum-independent-set or zero-forcing optimization contracts. Producer output remains COMPUTED; only the authorized independent replay is VERIFIED.

Validation

• focused graph producer/verifier and checker-installation suite: 25 passed
• Ruff lint and format: passed
• mypy on affected source: passed
• git diff --check: passed
• hosted CI will run broad lanes selected for shared graph contracts/checkers

Related to #947; does not close it.

[morluto]

Second-pass architecture review: rebase + likely merge. This is a clean atomic deterministic codec capability with explicit format rejection and an independent replay. It does not bundle the motivating NP-hard invariant work into the codec and stays within the domain-owned capability model. Existing issue #947 already tracks the broader graph6/invariant follow-up, so I would not open a duplicate issue for this PR.

[morluto]

Post-cutover architecture audit

Disposition: retain the mathematical slice, rewrite around jacobian.math.

A bounded exact graph6 decoder is a useful atomic codec operation and is not superseded by the cutover. Move decoding/canonicalization into a domain-owned graph encoding module and return a strict graph value whose vertex order/labels and edges are structurally bound; do not make the digest a substitute for that identity.

Install the producer as a thin OperationSpec and the standard-library replay as a separate authorized checker. Remove old graph bundle/exact-domain dispatcher/result metadata, relationships, and generic assurance fields. Keep the rejection cases for sparse6/digraph6/extended headers/padding and the H24 regression.

[morluto]

Post-#1265 review: keep the graph6 decoding/checking slice, but rebase it as a domain-owned graph conversion and remove unrelated stack residue. The operation is still absent from current main.

Graph6 is a mathematical/domain encoding, so an explicit graph-owned conversion is compatible with the current rule against generic codec frameworks. It should live with graph semantics (or a private graph encoding adapter used by the public graph function), produce the canonical typed graph value, and not create a universal serialization/coercion layer.

Keep the independent checker as a separate operation if it provides value. Do not carry old relationship metadata, and drop the unrelated discrete-log edits from the current branch when restacking.

Recommendation: revise/rebase rather than close. Preserve exact decoding/canonicalization attack tests and ensure malformed/oversized encodings fail before graph backend work.

[morluto]

Superseded by #1284. The replacement moves graph6 value/decoding semantics into jacobian.math.graphs, exposes one thin OperationSpec, and uses a dedicated standard-library checker runtime. It deliberately omits the old invariant mini-framework and reciprocal relationship machinery.