Assert every published argument pattern has a matcher

[lamemustafa]

Outcome and reason

Rebuilt on master after #444. #444 independently admitted amends_batch_id through tools/call, which fixed the same bug this PR first targeted. That bug: every amendment over MCP was refused as argument_invalid:amends_batch_id, found by the live lab end-to-end run described below. This PR now keeps only what #444 does not cover: nothing stops the next published pattern from failing the same way.

validate_string_bounds refuses every value published under a pattern that published_pattern_matches does not implement. amends_batch_id shipped like that, and 16 tests and two reviews passed because the tests called the handler directly.

Change

• agent_catalog.rs: published_pattern_matcher(pattern) -> Option<fn(&str) -> bool> is split out of published_pattern_matches, which now calls it. None means an unrecognised pattern. The three older arms keep their exact logic; the date arm moved verbatim into date_wire_matches. The batch-id arm now applies agent_import::valid_batch_id, the rule the build itself enforces, in place of Parse password-protected bank statements into voucher proposals, through PDFium #444's strip_prefix("bridge-").is_some_and(is_uuid_v4_lowercase). BRIDGE_BATCH_ID_PATTERN stays pub(super) for the schema, and is_uuid_v4_lowercase stays for proposals_id.
• agent_admission_tests.rs:
  • every_pattern_admission_reads_is_one_it_recognizes walks every registered tool's top-level string properties and top-level string array items, which is exactly what validate_tool_arguments reads. It asserts published_pattern_matcher has an arm for each pattern, asking the matcher itself rather than a list kept in the test.
  • the_batch_id_pattern_admits_exactly_what_the_build_admits refuses uppercase, a missing prefix, version 1, variant c, no hyphens, braces, a urn:uuid: form, the nil UUID, a trailing newline and a non-ASCII character. It also asserts valid_batch_id and strip_prefix + is_uuid_v4_lowercase agree on all those inputs plus 256 fresh v4 ids.
• Surface resealed; the independent seal check passes with 251 pins, 0 mismatches, and no pins added or dropped vs master.

Scope limit: the walk covers top-level patterns only. Nested patterns are read by validate_against_schema for voucher_presence, whose inventory test already pins them, and by the import payload's typed parser.

Validation and evidence

• Mutation: with the batch-id arm removed, every_pattern_admission_reads_is_one_it_recognizes, the_batch_id_pattern_admits_exactly_what_the_build_admits and Parse password-protected bank statements into voucher proposals, through PDFium #444's an_amendment_is_admitted_through_tools_call_argument_validation all fail. Restored, cargo test --locked --lib -- agent::catalog an_amendment_is_admitted_through_tools_call an_unknown_amendment_target passes 11 tests: the 9 in agent::catalog, Parse password-protected bank statements into voucher proposals, through PDFium #444's tools/call amendment test, and an_unknown_amendment_target_is_refused_before_any_tally_read. Without the last filter, 10 pass. cargo fmt is clean, and cargo clippy -p bridge --lib --tests -- -D warnings -A clippy::pedantic is clean.
• Live (licensed 7.1 Silver, synthetic BRIDGE SHAPE LAB, first version of this PR): MCP build_import_xml with amends_batch_id was refused at master f25fc91a. With the fix it was admitted; the compare-and-swap matched both vouchers; the amendment imported ALTERED=2 and verified 2/2; and a later change behind Bridge's back was refused with book_voucher_diverged. The admission rule was the same valid_batch_id used here.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.