fix(cli): require a loadable bank for each run locale before any model call

[3mei]

What changed and why

The asset pre-flight checked only that a locale's folder exists. A folder can exist and still fail every row.

• The pre-flight loads each locale-partitioned bank (financial_services, sov_ai_facts, sov_ai_dynamic) for the run's locales, after the India-variant mapping, with the loader the probe itself uses. A refusal names the locale, the reason (no bank, or the loader's error) and the locales that can run.
• Every path runs it. simulate --dry-run also checks a panel's locales; --materialize-inputs and --inputs run it, the latter checking each probe against the locales its own rows use, under the asset root they run on; notebook 01 calls it before its run loop.
• Docs. architecture/probes.md describes the check, and notebook 01's comment no longer lists hi_Deva_IN as a finance locale.

How it was verified

• make check-all: lint clean, 333 files already formatted.
• make test: 4087 passed, 2 failed. The two failures are in tests/test_replay_reasoning_config.py, which read a models.local.toml when the checkout has one; they pass without it.
• make smoke, make check-license-headers, make check-doc-links, make check-notebooks-clean and make check-dist pass.
• New tests use a fixture asset tree (a copy of the en_US finance bank, plus an fr_FR copy without tasks), so they don't depend on which banks ship. They cover a bank that doesn't load, the locales that can run, the dry runs with --locale and --panel, --materialize-inputs, and --inputs checked probe by probe. Each fails without the change.
• On the shipped banks: simulate --dry-run --locale hi_Deva_IN --probe-mix financial_services=1 exits 2 with "Locales that can run: en_IN, en_US"; ta_Taml_IN runs through en_IN; notebook 01's default settings pass.
• No make health-check: no change to models, providers, prompts, scorers or the trajectory store.