refactor: stabilise bench-tx.json keys and consolidate cost-table tests

[mmagician]

Stacked on #3714. Review the top four commits.

PRs touching MASM or kernel code carry large generated diffs, and bench-tx.json's share of them is mostly noise. Its note_execution map was keyed by a note commitment hex and stored in a BTreeMap, so any change to an inlined note script renamed every key and re-sorted the section.

In #3699 both P2ID notes in the two-note scenario moved by exactly +22 cycles (2109→2131, 2324→2346), but because the new hexes sorted the other way the diff read as 2109 → 2346 (+11%) and 2324 → 2131 (−8%). A reviewer scanning line by line drew the opposite conclusion from the truth.

What changes

note_execution becomes an ordered array keyed by note kind:

"note_execution": [
  { "note": "P2ID#0", "cycles": 2131 },
  { "note": "P2ID#1", "cycles": 2346 }
]

Labels come from the note's script root, so they move only when the note's kind, its multiplicity in the transaction, or its position among same-kind notes changes. Array order is the order the kernel measured the notes, so it never re-sorts.

Each scenario also gains total_cycles - the figure the cost tables are derived from - so the snapshot carries it alongside the per-stage counts.

Cost tables are deliberately untouched

Regenerating would move every constant, but every move is within the 5% drift tolerance (worst: BURN at −3.06%), so the guard never asked for it. A refactor has no business shifting fee-affecting values. One consequence worth knowing: make update-note-costs is therefore not a no-op on this branch — it rewrites the tables with that accumulated sub-threshold drift. That is pre-existing on next and is the churn the follow-up PR's policy addresses.

A bug this surfaced

The join needed real note IDs and did not get them: TransactionMeasurements::note_execution is typed Vec<(NoteId, usize)> but carries each note's details commitment, because the host reads the input note segment's base word instead of INPUT_NOTE_ID_OFFSET. Fixed separately in #3724, based on next.

This PR does not depend on that fix. It joins on the details commitment, isolated in measured_note_key with a TODO stating the premise, linking #3724, and warning that the migration is all-or-nothing — every lookup misses at once, so whoever lands #3724 must update this in the same PR. A details commitment is weaker than a note ID (it excludes the metadata the nullifier binds), so duplicates are rejected rather than silently misattributed.

Test consolidation

checked_in_cost_matches_benched_cycles and created_notes_cover_executed_output_notes each walked PricedNote::all and executed ~31 scenarios independently. Merged into checked_in_note_costs_match_executed_scenarios, which executes each once and asserts both properties: ~62 transaction executions per test run become ~31. The two failure kinds are still reported separately. That pass also validates the label join, which nothing else exercises against a real transaction.

.gitattributes

bench-tx.json and Cargo.lock are marked linguist-generated=true. Limited to files something automated already checks, since collapsing a diff takes a reviewer's eyes off it. The cost tables and the agglayer Solidity test vectors are deliberately excluded — the latter are bridge conformance oracles regenerated only by a manual make generate-solidity-test-vectors, with no CI job watching them.

Verification

• make test-release: 1886 passed
• make lint clean
• Regenerated twice; the two runs differ only in range_rows on 7 scenarios, the run-to-run variance the README documents. Labels, structure and all cycle counts identical.
• No cost table constant changed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LXcU9HrrmWTCTZTJZnuLSU