feat(pi): add validated MLflow OSS streaming

[dgokeeffe]

Summary

• add Pi's validated MLflow/OSS provider and capability metadata
• repair incomplete SSE termination from compatible OSS backends
• restore the direct gateway URL and close the loopback proxy on normal exit, errors, SIGTERM and SIGHUP
• enforce managed Pi model allowlists without falling back to unlisted discovery results
• preserve model/session affinity and supplemental capability state

Why

Pi could not reliably use coding-oriented OSS models through AI Gateway. Some backends omit the terminal stream fields Pi expects, and an interrupted session could leave the persisted model configuration pointing at a dead loopback port. This adds the compatibility layer with bounded lifecycle and policy handling.

Stack

PR D of five. Depends on PR C (pr/capability-discovery). It is opened as a draft until A, B and C land; the isolated review delta is 763fa3e..5ac7937.

D and E are siblings and can be reviewed or landed independently once C is available.

• A: #446 — Pi routing foundation
• B: #447 — Claude discovery hardening
• C: #448 — capability-driven discovery
• D: #449 — this PR
• E: #450 — sibling OpenCode PR

Validation

• 99 final focused Pi/proxy tests passed after dependency replay
• SSE termination, truncation, signal handling, cleanup and managed-allowlist cases are covered
• Ruff lint and changed-file formatting passed
• ty check src and git diff --check passed
• fresh installed-build Pi inference passed on system.ai.deepseek-v4-pro-0813
• the isolated proxy URL was restored and its listener was closed after inference
• the complete integration stack passed 2,358 tests with 37 skips; its three failures reproduce unchanged on main

Residual risk

An uncatchable SIGKILL or power loss can still interrupt cleanup. A subsequent Pi launch repairs a stale loopback URL before starting a new proxy.