report each compared node's trace in the graph matrix, on request (ADR-0031)

[kikashy]

Closes #127.

ADR-0031: experimental_test_graphs accepts an optional boolean include_traces (decoded exactly as ADR-0028's rehearsal boolean — exact spelling, JSON boolean only); experimental graph test grows --include-traces on both forms. Asked, each reported comparison of an evaluated node carries the evaluation's own ADR-0027 trace, by pointer so presence tracks the request (asked is present, [] at minimum; not asked is absent — off is the previous payload byte for byte). An undeclared node was never evaluated and carries none; a row that fails before node comparisons carries none. Traces ride inside each row's marshaled report, so the existing ADR-0026 budget charges them with no new machinery.

ADR-0027 §0's matrix-results-carry-no-traces determination is partially superseded for graph matrix node comparisons via the index annotation; its text is untouched, per the ADR README's partial-supersession rule.

Tests assert the wire: presence on every compared node across MCP/CLI/graph layers, absence (not emptiness) when unasked with declined-equals-omitted byte identity, trace retention on a mismatching comparison, no trace on an undeclared node, byte equality between the comparison's trace and the evaluate composite's for the same node and inputs, and case-fold/type refusals for the new member. Six mutations each fail a test; the one safeguard that cannot discriminate is the nil-floor guard in the attach (unreachable while the evaluator honors ADR-0027 §1's never-nil floor), labeled here as such.

Material-decision impact: public-surface; review: round to be posted on this PR

🤖 Generated with Claude Code

[kikashy]

Review round 1 — disposition

Reviewer: Codex (codex-cli 0.145.0, OpenAI) · Reviewed SHA: d2cd53b9315d1747e1b9c48218a556bd3c6aa959 · Fix commit: 74bbbed (covered by the dispositioned-finding exception; 34e2e93 is the status flip only)

#	Severity	Finding	Disposition
1	MEDIUM	Traced budget behavior not mutation-tested: no test where a bare run fits and a traced run is refused.	Accepted, fixed. Differential boundaries on both paths with no magic numbers: the reconcile makes the bare spend exactly the envelope's (direct) / entry's (walk) own bytes, so the bare run passes at that exact budget and the traced run is refused with JPS-GRAPH-REPORT-BUDGET. Mutation-verified: disabling all three charge sites in rows.go fails the test.
2	MEDIUM	Public-interface gaps: CLI walk wiring untested, MCP refusals asserted by isError only, tools-list doesn't pin the new schema member.	Accepted, fixed. Walk form asserted like the path form (deleting its wiring fails); a rehearsal-style full-message test pins every refusal verbatim including number/object/array values and both case-folded spellings; the tools-list pins the closed schema's optional boolean include_traces with no required members (type drift to string fails it).
3	LOW	Pointer/shape edges unpinned: two-node row, empty-trace-as-[], headline-mismatch-with-expectedNodes.	Accepted with one declined part. The two-node row now pins each comparison to its own node's trace (byte equality against a direct evaluate, plus an inequality guard; the borrow-another-node's-trace mutation fails it), and a headline mismatch is pinned to zero comparisons and zero traces. The empty-trace transport case is declined: no fixture pack this evaluator admits produces an empty trace, so the case would require inventing an unreachable shape — the nil-floor guard covering it is defensive, cannot discriminate, and is labeled as such in the PR body per the repo's mutation-check practice.
4	LOW	--include-traces invisible in human format while help implies otherwise.	Accepted, fixed (intended branch). Human rendering is deliberately unchanged; flag help, Long text, the ADR (new Neutral consequence), and the changelog now all say the JSON report carries it.
5	LOW	Two stale two-member comments.	Accepted, fixed. testGraphsArguments doc says all three members; the refusal-table comment names the optional boolean.

Also adopted from the verification notes: the declined-equals-omitted equality now compares literal wire bytes (toolText), not a remarshal that could hide an unmodelled member.

Reviewer's sandbox could not start go test (read-only FS); the full gate ran author-side at 74bbbed: gofmt/vet clean, full suite green, corpus 20/20. Nine hard-verification items (off-output byte identity, pointer aliasing, budget escape, supersession mechanics, decode parity, CLI wiring, mutation table, roster/claims, ADR conventions) recorded in the review log; no functional runtime defect was confirmed.

ADR-0031 flipped to accepted (34e2e93).

[github-actions]

🎉 Thank you for your contribution to Judgment Pack!

If you’d like another task, choose a path that interests you:

• Go runtime
• Gateway and provenance
• Python and evaluator experiments
• Demos and applications
• Specification and documentation