Stop the examples spending money, and gate README flag drift

[kmadan]

Both of these are things a first-time contributor hits before writing a line, found while checking that #51 is actually startable.

The examples bill you for running them

All three examples/*/run.py call OpenAI. DeepEval judges toxicity with an LLM and activates off the presence of OPENAI_API_KEY, so:

• anyone with a key in their shell gets billed for running an example
• anyone with an exhausted quota watches nine retries per interaction scroll past and reasonably concludes the example is broken

The bundled demo already guards against exactly this, popping the key unless --with-llm-metrics is passed. The examples never got the same treatment.

before:  ERROR:root:OpenAI Error: 429 ... Retrying: 9 time(s)...   (×many)
after:   0 OpenAI calls, same report, 54s

Opt in with AICERTIFY_WITH_LLM_METRICS=1, documented in each example README.

The guard must run before from aicertify import ..., because the evaluator stack reads the key at import time. That trips ruff's E402, so the import carries a noqa and a comment explaining that the ordering is load-bearing. Reordering to satisfy the linter would reintroduce the bug.

The README flag tables had drifted in both directions

Flag	State
--evaluators	removed in #91, still documented in all four translations
--params	has always existed, documented in none of the five READMEs

Neither shows up in a test, a lint or a build. A reader following the Japanese README gets unrecognized arguments and concludes the tool is broken.

tests/test_readme_flags.py now asserts every flag any README documents actually exists, across all five languages, and that the English table covers every evaluate flag. It introspects the argparse tree rather than scraping --help, and has a guard-the-guard test so introspection breaking cannot make the others pass vacuously.

Mutation-tested before committing: removing --params from the English table fails; putting --evaluators back in the Japanese one fails. Restored, green.

The assertion is scoped to the evaluate subcommand rather than allow-listing exemptions. An allow-list of "documented elsewhere" flags is somewhere to quietly park anything inconvenient, and it grows until the check means nothing. It did immediately surface four more undocumented flags on demo and init-contract, which are described in prose.

The translations still lack a --params row. That needs a native speaker rather than a machine translation nobody can vouch for, so the test does not demand it and it stays a contribution task.

268 tests pass.