feat(pi): add validated GLM/Kimi MLflow provider

[dgokeeffe]

Summary

Adds Pi support for the validated OSS coding-model cohort already used by ucode: GLM and Kimi.

Closes #204. Supersedes the earlier broad-cohort/proxy approach in this PR.

Changes

• Add Pi's databricks-mlflow provider (openai-completions → /ai-gateway/mlflow/v1).
• Configure the gateway compatibility flags Pi needs: supportsStore: false and supportsStrictMode: false.
• Share GLM/Kimi context, output, and reasoning metadata across Pi and OpenCode.
• Let Pi select discovered OSS models and use them as the final default fallback.
• Keep OSS discovery as an explicit product allowlist: GLM and Kimi only.
• Keep the independent gpt-oss exclusion from the Codex/Responses bucket so it cannot be routed to an incompatible API.
• Include Pi in OSS discovery diagnostics.

Review follow-up

• Removed the local SSE-repair proxy entirely. The missing finish_reason is an AI Gateway defect tracked by AI Gateway MLflow chat-completions omits finish_reason on natural stop (breaks inkling in Pi) #215 and should not add an HTTP proxy to ucode's launch path.
• Inkling is deferred until AI Gateway MLflow chat-completions omits finish_reason on natural stop (breaks inkling in Pi) #215 is fixed. It is not offered to Pi or OpenCode by this PR.
• Llama, Qwen, GPT-OSS, Gemma, and other chat models remain unlisted until they are explicitly validated for coding harnesses and accepted into ucode's supported model policy.
• Removed the unrelated version bump and uv.lock changes.

Validation

• uv run pytest tests/test_agent_pi.py tests/test_agent_opencode.py tests/test_databricks.py tests/test_cli.py -q
• uv run ruff check on all changed Python files
• Final review-loop validation: 588 passed, 29 skipped across the focused changed-area suite.
• The full stacked branch passes 1062 passed, 36 skipped when installed-agent capture tests are excluded. Those tests depend on locally installed agent state rather than the unit under review.

Scope

This is the base of the stack:

1. feat(pi): add validated GLM/Kimi MLflow provider #217 — Pi GLM/Kimi provider and shared OSS metadata
2. fix: declare long-context Claude and GPT metadata #223 — Claude/GPT long-context metadata
3. fix(opencode): add GPT routing alongside OSS discovery #239 — OpenCode GPT routing and serving-endpoint fallback

[dgokeeffe]

@rohita5l @AarushiShah-db — would appreciate a review when you have a moment. This builds directly on your OSS work (#178 Kimi, #186 GLM): adds a Pi databricks-mlflow provider (Pi had none), broadens the OSS family bucket to the rest of the chat-completions cohort (inkling/llama/qwen/gpt-oss/gemma, excluding embeddings), corrects the GLM limits to the probed gateway caps (1M context / 65536 output), and adds a reasoning flag. One heads-up: it reverses the GLM output cap from #186's 25k to a verified 65536 — details in the PR description. A code-review pass was already applied (proxy framing/error-handling/SSRF-guard hardening). CI hasn't run — could use a maintainer approve-and-run on this fork PR.

[dgokeeffe]

@rohita5l I refreshed this PR around the review feedback: removed the local proxy, deferred Inkling pending #215, narrowed OSS selection to GLM/Kimi, removed version/lockfile noise, and resolved both threads with details. The branch is rebased on current main and locally validated. Could you re-review and approve the fork CI run when convenient?

[dgokeeffe]

Post-review update: a 3-round fresh-context review loop found and fixed two additional foundation issues in this layer: GLM 5.2 now has model-specific 1M/65536 limits while older/unknown GLM keeps 200k/25k, and Pi OSS-only workspaces now pass shared availability/diagnostic checks. Focused and full-stack validation are updated in the PR body.