Self-hosted model: a deploy Modal never schedules waits silently; Claude Code settings keep the caller's session variables #33

Found on Nov 11 while Greenswarm started a run on a self-hosted Qwen through FACTORY 0.7.1 (Greenswarm issues flujo-app/greenswarm#318 and #324).

1. A deploy that Modal never schedules waits silently

On a Modal "starter" plan a function with a request time limit of 45 minutes or more is never given an H100. The qwen profile asks for 60 (request_timeout_minutes), nemotron too. Modal says so in the app log ("... is waiting to be scheduled on a GPU_H100 worker. Your workspace's concurrency limits are reached (H100 unavailable on starter plan)"), but ModalSelfHostedModel.deploy() only sees a held request and waits up to 45 minutes per replica, then reports an uncertain outcome. Measured with test functions of the same size: 10 to 30 minutes start in 3 to 4 seconds, 45 and 60 minutes never start.

Wanted: while verifying, read the app's log for the scheduler's "waiting to be scheduled" line and stop with that text; and a default request limit in the profiles that the common plans accept (30 minutes worked).

2. Claude Code settings for a self-hosted provider leave the caller's Claude session variables in place

claudeProviderEnvironment() sets the base address and the key and clears ANTHROPIC_API_KEY and CLAUDE_CODE_OAUTH_TOKEN. When the caller itself runs inside a Claude session (a desktop or SDK session), the child also inherits that session's variables (CLAUDE_CODE_SDK_HAS_HOST_AUTH_REFRESH, CLAUDE_CODE_ENTRYPOINT, CLAUDECODE, USE_STAGING_OAUTH, ...). Claude Code 2.1.321 then offers the session's own login to the endpoint instead of the endpoint key: HTTP 401, and a login is sent to a server it was not meant for. An older Claude Code (2.1.289) hangs instead.

Greenswarm now clears every inherited variable whose name starts with CLAUDE or ANTHROPIC_ or contains OAUTH before applying FACTORY's settings (swarm/lib/endpoint.mjs, sessionVariablesCleared). Wanted here: the same in claudeProviderEnvironment() (null for those names), so every caller gets it.

3. Small: a stand-in endpoint for tests has to speak https

validateModelProvider() refuses http://127.0.0.1:<port>. That is right for real endpoints; for tests a loopback address could be allowed. Greenswarm's stand-in makes its own certificate meanwhile.