Capture real protocol evidence for aiosendspin's server-initiated-protocol-baseline-v1

[chrisuthe]

What

Implements real wire/state evidence capture (CORE-001..004, PLAYER-001) for the \server-initiated-protocol-baseline-v1\ scenario in both aiosendspin adapters, replacing the placeholder scaffolding merged in #109.

Approach (interim, tracked in aiosendspin#111)

aiosendspin has no first-class tracing hook yet, so this uses a fragile-but-honest introspection approach, decided on explicitly with the user:

• Server side: runs the harness's \SendspinServer\ with \�llow_noncompliant_clients=False\ only for this scenario (all other scenarios keep the permissive default). The server SDK already enforces early-message rejection, role-object gating, and the initial \client/state\ gate internally via \lag_noncompliance()/\ClientComplianceError\ — a successful connection under strict mode is itself proof CORE-002/003/004 passed, without reimplementing spec validation in the harness. \client.send_message/\send_role_message/\send_binary\ are wrapped (same pattern as the existing PCM scenario code) to observe the negotiated \stream/start\ format and audio chunk timestamps for PLAYER-001. Handshake start/end timestamps are captured bracketing the adapter's own await calls for coarse CORE-001 evidence, plus post-hoc \handshake_hash/\psk_category\ read from the connection.
• Client side: since \�ttach_websocket()\ blocks until disconnect, the live \SendspinConnection\ is captured while it's still running (not after) so CORE-001 evidence isn't lost. Reuses the client's existing stream/audio-chunk listeners to independently record CORE-002/003/004/PLAYER-001 evidence from the client's own vantage point — required because \protocol_evidence_failure()\ checks both roles' summaries independently.

Verification

• \python -m compileall src scripts\ passes.
• Local matrix smoke test (aiosendspin -> aiosendspin): \server-initiated-protocol-baseline-v1\ now passes all 5 protocol assertions; the rest of the existing scenario matrix is unaffected (same pass/fail pattern as baseline, aside from one pre-existing timing-flaky metadata case unrelated to this change).

Follow-ups

• Replace this introspection approach once aiosendspin exposes first-class tracing hooks (aiosendspin#111).
• Extend other implementations' adapters to support this scenario in follow-up PRs.