[jithinAB]

A runtime that runs PACT one operation at a time has to know which fields reach each of the 76 operations of the seven primitives (97 §2, §5.4). That list lived in one place, the OPS table inside scripts/pattern-coverage.py, and said it in keyword prose (limits, at-most, fuel, checks-at-most), which a runtime can only turn into fields by matching words. Bud Meta Harness did exactly that for its per-operation registry and got it wrong in both directions: shape.stream picked up bundle.from and provenance.harness, and shape.one-of picked up nothing. This puts the operations in a file and puts on each field the operations it is, so the mapping is data both sides read, held to each other by a test.

Nothing changes in what an author writes or what pact check accepts.

What changes

• spec/kernel.yaml: the 76 operations, each with its primitive, what it does and the keyword that is it today, copied from OPS unchanged. An operation an author writes as syntax inside a value rather than as a field of its own says where with written-in: (six of them: one of a, b, list of, stream of, an agent-shaped answer, a binding, a list of conditions).
• kernel-op: on a field of spec/schema.yaml: one operation or a list. 291 field lines, 295 fields once the anchored interface:/chosen-by: are counted on skill and tool, 347 field-operation pairs. Descriptions, names and because: carry none. Placed after each field's tier: line, so no run of type:/names:/key-names:/surface: lines that a test anchors on is split. The header's attribute list describes it.
• Field::kernel_ops in pact-schema, read by from_doc with the same helper as names:.
• crates/pact-schema/tests/every_kernel_operation_is_written_somewhere.rs (4 tests): every operation names one of the seven primitives, is named <prefix>.<what>, and says what it does and its keyword; every kernel-op: names an operation the kernel lists; every operation is named by a field or placed by written-in:; the attribute travels through an anchor. The last test runs the same check on the real files with one line changed each way (store.fork renamed on stage.forks, and shape.one-of's written-in: removed) and requires both to fail with the sentence that names them.
• scripts/pattern-coverage.py reads OPS from spec/kernel.yaml; its output is byte-identical to master's. --ops gains a column with each operation's fields.
• scripts/primitive-coverage.py adds the primitive of each operation a keyword's kernel-op: names after the primitives its table gives it. The main ingredient never changes. 48 keywords gain a secondary primitive where the table had given them fewer than their operations have; agent.accepts is actor.signature by OPS's own keyword list, so it is an Actor as well as a Shape. docs/97 §6's "any part" column, Appendix A (48 rows, each only appended to) and Appendix C (the new column) are regenerated; the §5.4 sentence points at the file.
• CHANGELOG entry; headline test counts synced.

Evidence

• cargo test --workspace --locked: 1652 passed, 0 failed (the headline count is synced to it).
• cargo clippy --workspace --all-targets --locked -- -D warnings: clean.
• python3 scripts/pattern-coverage.py: same output as master; --ops lists fields for 75 operations and none for shape.one-of, which is placed by written-in:.
• python3 scripts/primitive-coverage.py: 569 keywords, 0 unmapped, main-ingredient counts unchanged.
• The full adapter suite: 3169 passed, 151 skipped, 1 failed: test_the_second_port_reports_a_spend_cap_it_cannot_price, which fails on master too on a machine without the TypeScript adapter's dependencies installed.

Choices a reviewer may want to change

• The mapping is mine, curated field by field against OPS's keyword column and the help text. Where a field is several operations it says so (stage.does is six; when-this.in-share is shape.digest and check.digest, as OPS lists it under both). Every placement is a line of YAML to move.
• written-in: is a sentence, not a list of types, because the six places are syntax a reader recognises, not fields a checker can resolve.
• The floor in the anchor test (more than 340 pairs) is there so the list cannot shrink silently, the way spec/comparisons.yaml's row floor works.

Downstream: Bud Meta Harness reads kernel-op: at its next re-pin instead of matching keyword text (its PK-1 kernel registry and coverage dashboard).

🤖 Generated with Claude Code

[jithinAB]

Merged master (0299680, after #29) into this branch. CI on the merged head: gate / everything passes. gate / newest-pydantic-ai fails only the 5 tests/test_parity.py cases master itself fails (pydantic-ai's newest release adds capabilities the registry does not list yet), as on master's run 43389520212. Marking ready and merging, per the owner's standing rule for pact-kernel's upstream PRs.