TOOL_PARSE_CORPUS

The tool_parse_* conformance corpus

The behavioral corpus for the text tool-call parser, migrated from ~4900 lines of Rust unit tests into executable conformance fixtures for the harn#5142 parse re-platform. Every baseline was generated from the observed stdout of the built binary, never from reasoning and never from the Rust assertions, so the corpus pins what the parser does rather than what anyone believes it does.

The port's contract is: make these pass unchanged. A diff in any .expected here is a behavior change and needs an argument, per conformance/README.md.

Layout

_tool_parse_probe.harn is a helper module, not a test — it has no baseline file, so the runner skips it the same way it skips _common.harn and the modules/*_lib.harn helpers. It owns the four registries the Rust corpus used (sample_registry, corpus_registry, echo_registry, native_registry) and the printers every fixture reports through.

Keep the registries' parameter types intact when editing them. They are not decoration: chat-template markup parses a parameter value as JSON when the schema is non-string and keeps it verbatim when the schema is a string, so content: {type: "string"} versus ops: {type: "list"} is the difference between two fixtures passing and failing.

What the common record prints, and what it does not

report prints, per case: call count, per-call name plus a key-sorted json_stringify of the arguments, every parse error and protocol violation in full, the stray-recovery count, prose, user_response, and done_marker.

Three observable facts are deliberately excluded from that record and live in dedicated fixtures instead, because they are expected to move independently of dialect behavior. Keeping them out means a port that re-cuts one of them diffs one file rather than all twenty-six.

Excluded fact	Fixture	Why it is separate
dropped fragment granularity	tool_parse_drops	Whole-unit fragments instead of line slices are fine. The invariant is that something is recorded, not where the cuts fall.
canonical_text	tool_parse_canonical_replay	Long, and orthogonal to whether a call parsed.
Synthetic call ids	tool_parse_turn_unique_ids	The counter is process-global. The Rust test asserts tc_0/tc_1; run under the full suite the same input yields tc_265/tc_266. Pinning literals would pin fixture execution order, so that fixture asserts the contract — consecutive within a turn, no reuse across turns — instead.