feat(conformance): wave 15 — final VM sweep, every real API adapter tested

[deblasis]

The campaign's closing wave: engine-level suites for all 21 remaining real-API adapters — 235 new behavior markers, 1,079 VM behaviors total. VM tier 38 → 59, boot tier 24 → 3 (the three 0-route demos only). Every real API adapter now carries a real test suite.

~25 adapter fidelity fixes, the standouts:

• avalara: _to_float returned 0.0 for all strings — the adapter's own documented SDK decimal-string requests computed zero tax
• pinata: CIDs are now real CIDv0 multihashes (identical content dedupes with isDuplicate: true); file pins actually read the multipart upload; real default pageLimit 10
• cloudkit: every record delete 500'd (the store was handed a dict where it wants the id); a JSON-number resultsLimit 500'd on len()
• tenderly: _hex_pad coerced 0 → 1 — every simulation shared one transaction hash and zero-value traces rendered 0x1
• linkedin: post resolution returned the caller as author; float-in-URN ids (urn:li:share:1.0); OAuth code burned before client validation (the fifth instance — the standardization sweep is now complete); forged comment actors accepted
• dropbox: re-upload forked the path (stale reads forever) — and the conflict fix now honors the real write-mode contract (overwrite/update replace in place, autorename forks a suffixed path, default add conflicts)
• persona: the one-shot terminal webhook was consumed on the created→pending hop — polling clients never received inquiry.completed
• revenuecat dual-key model; walletconnect symKey + relay echo; sendgrid half-configured webhook rejection; etherscan NOTOK envelope; opensea V1ErrorWrapper + bytes32 zone_hash; onfido blank-field 422s; signin-with-apple needed zero fixes — its ES256/JWKS round-trips verify cryptographically

36 new sidecar bullets (reviewer-verified against handlers); duplicate revenuecat/sendgrid bullets and a wrong etherscan missing bullet corrected; two engine-test response-body leaks fixed. Reviewed by code-reviewer agent: no critical findings; all five importants fixed (including the dropbox mode regression the reviewer caught in my own conflict fix).