fix: recognize currency-formatted numeric policy arguments

[fatihcvs]

Summary

Fixes #66. Numeric inequalities now accept a leading $ and valid comma-separated thousands groups, so $250 and 1,000.50 reach the same policy boundary as JSON numbers. Validate grouping before removing separators, so malformed values such as 1,00 remain non-numeric. Equality and $context. references are unchanged.

Document the contract in SYSTEM_DESIGN and cover all four inequality operators, malformed formats, context/equality behavior and actual refund-policy violations.

Type of change

• Bug fix
• Tests
• Core engine (policy, campaign, minimize, regression)

Checklist

• Read CONTRIBUTING.md
• Linked related issue
• Core remains free of framework SDK imports
• Behavioral tests added and unit suite run
• Documentation matches the implemented behavior
• One concern per PR
• Authorized local testing only

Test plan

• New policy tests against unchanged code: 26 failed, 133 passed.
• pytest tests/unit tests/integration tests/reliability -q --tb=short: 555 passed, one existing smoke script failed printing a Unicode character under the Windows default encoding.
• With PYTHONUTF8=1, pytest tests/reliability/test_smoke_gate.py -q --tb=short: both tests passed, including that failing case.
• Ran with MUTINY_SAMPLE_OFFLINE=1, isolated database/cache/temp paths, and uv available on PATH (the initial unit run could not locate uv for its wheel-build test).
• git diff --check passed. No live model or hosted service was used.

[vercel]

@fatihcvs is attempting to deploy a commit to the priyanshu's projects Team on Vercel.

A member of the Team first needs to authorize it.

[CodewithJha]

Thanks @fatihcvs — this hits #66 / MUT-017 cleanly: _as_number now accepts a leading $ and validated thousands groups ($250, 1,000.50) for inequality operators, while $context. refs, equality, and malformed grouping (1,00, 250$) stay fail-closed. The unit matrix covers all four inequalities plus the refund_limit path, and SYSTEM_DESIGN documents the contract. GitHub Actions green on 3.11/3.12; ignoring Vercel auth.

[CodewithJha]

Merged — thank you @fatihcvs. Exactly the kind of focused Core policy fix Mutiny needs: LLM-style "$250" / "1,000.50" no longer slip past inequality constraints, malformed grouping stays fail-closed, and $context. / equality are left alone.

If you'd like to continue working on Mutiny, these would be good related areas to look at:

• #1 — policy operator cheatsheet; the $ / thousands-group contract you just documented in SYSTEM_DESIGN belongs there too
• #65 — another Core correctness hole (use_boundary_seeds / elite_count) in the same policy-campaign neighborhood
• #7 — small policy.yaml comment-header polish if you want a lighter docs bite

No pressure — only if you want another one.