[VickyXAI]

Port of the TypeScript SDK's #92 (from the review of its skip-booking change, which #94 here mirrors).

Problem

tx_log.settlement_skipped() reads any X-Settlement-Skipped value, or X-Payment-Waived when present, as "no money moved", and every booking site then books $0. Two gaps:

1. A marker can hide a real charge. If a response carries a skip marker and a receipt proving settlement (a transaction or a positive charge), the session total, the cost log and the spend limit all record $0 for money that moved. The gateways never send both today, but nothing pinned that.
2. Opposite values count as a skip. X-Settlement-Skipped: false (or 0, no, none) was a skip.

Change

• settlement_skipped() ignores a marker whose value is false / 0 / no / none.
• A marker is also ignored when the response proves settlement: a PAYMENT-RESPONSE that did not fail (success is not false) and names a transaction (via decode_settlement_header, which reads transaction / signature / …) or a positive amount / extra.chargedAmount, or a non-empty X-Payment-Receipt. The conflict is written to stderr and the call is booked.
• Gateway check: blockrun.ai's free-fallback answer is PAYMENT-RESPONSE {success: true, transaction: ""} with no amount and no X-Payment-Receipt (blockrun chat/completions/route.ts ~4742–4777). The pm empty-result and rpc unexecuted routes return before settling. So every skip sent today still books $0, and the real free-fallback shape is a test case.
• Kept as fix: book $0 for paid calls the gateway served without settling #94 designed it: an empty X-Payment-Waived still counts as a skip (presence is enough).

Tests

• tests/unit/test_settlement_skipped.py: these now book the quote: negative values; marker + receipt with a transaction / Solana signature / chargedAmount / amount; marker + X-Payment-Receipt. A conflict writes to stderr. A free-fallback receipt and a failed receipt stay skipped and write nothing.
• 8 fail on main. black and ruff are clean. pytest tests/unit: 1431 passed. Two test_solana_poll_ownership.py tests failed once under machine load (a different test each run), then passed 64/64 twice on rerun; they don't touch this code.