[3mei]

What changed and why

The asset pipeline now generates banks for markets that write numbers differently, and for institutions that don't offer every common operation. One commit per change:

• Amounts and rates in the locale's notation. Each typed product value now reaches the generation prompts with a display form in the locale's notation (7,90 € and 15,9 % in France, £7.90 in the UK), built by format_amount and format_percent, which the probe's money formatting now shares. The document prompt and the cluster judge ask for that form. Without it, about a fifth of a French test corpus copied English decimals (7.9) from the prompt's JSON.
• number_notation in validate-assets. A decimal-comma locale fails on any English decimal written before a unit, which the numeric-faithfulness check can't see once the right form appears somewhere.
• omit_common_operations. An institution can drop injected common operations it doesn't offer: a building-savings bank issues no cards. Naming an operation its type isn't given, or both declaring and omitting a tool, is rejected. SCHEMA.md documents it.
• Discoverable tool descriptions. A description in the region spec steers the generated document for a discoverable tool, while tools.yaml keeps it empty, so the assistant still learns the tool only by retrieving that document.
• Variable-name check. A one-word key such as excess is also the ordinary word, so the check matches only snake_case keys, as the tool-name check already does.
• Docs wording. The tools.yaml header gen-assets writes, the probe's docstring and README, and the spec comments describe the discoverable-tool gate by what it does: a tool is callable only after a document describing it has been retrieved. The 14 shipped tools.yaml headers match.
• Label test. The label-coverage test discovers every region spec that authors variable_labels, rather than listing locales.
• Scorecards stay local. evaluate-assets writes _eval_scorecard.parquet beside a bank. Like embeddings.parquet, it is now gitignored, excluded from package data, and rejected by make check-dist.

Each fix has a test that fails without it.

How it was verified

$ make check-all
All checks passed!
334 files already formatted
$ make test        # bundled models config, as CI runs it
4128 passed in 30.13s     (top of the stack; every commit in this PR also passed alone)
$ make smoke
All checks passed.
$ make check-license-headers && make check-doc-links && make check-notebooks-clean
🎉 All files have correct license headers!
✓ every relative link, and every link into this repository's main branch, resolves
OK 4 notebook(s) carry no outputs
$ make clean && make check-dist && make check-extensions
✓ installed package runs from a clean environment
all extension seams work against an installed distribution

The probe's money formatting now goes through format_amount, so the run path was checked live, against a custom provider config:

• make health-check with HEALTH_PROBE=financial_services HEALTH_LOCALE=en_US: simulate, eval and report completed, with no failed model requests.
• Scorer check: one 8-turn en_US conversation evaluated with --scorers financial_services. The judge scored it, and the scorer returned ok: ordering 1.0, document recall 0.83, no unauthorized action 1.0, substantive retrieval 0.71. Tool selection was 0.0: the task needed a contact update, a withdrawal request and an account closure, and the assistant searched the knowledge base 26 times without calling any of them.

The same code generated the fr_FR, de_DE and en_GB banks in the PRs stacked on this one, where validate-assets passes number_notation for every institution.

Results impact

None for the shipped banks: their text is unchanged except the tools.yaml header comment. New generations write numbers in the locale's notation, and validate-assets can now fail a decimal-comma bank that it passed before.