feat: allow spawn_agent model and reasoning effort override - #27385

Adds explicit model selection to subagent spawning. A parent agent can now discover which chat model configs are usable and spawn a child subagent on a specific model and reasoning effort, instead of always inheriting the configured default chain. The transcript shows the selection on the spawn row.

Changes

• New list_subagent_models chatd tool: returns the caller-usable model configs (UUID, display name, model, provider, context limit, default flag, selectable reasoning efforts). Hidden for delegated child chats, same gating as spawn_agent.
• New optional spawn_agent args:
  • model_config_id (UUID from list_subagent_models): the child chat is created on that config. Wins over personal and deployment overrides and over parent inheritance.
  • reasoning_effort (none/minimal/low/medium/high/xhigh/max): pins the child's effort, clamped to the selected model's supported range at generation time. Independent of model_config_id.
• Hard failures, never silent fallback: malformed or unknown UUID, disabled config, disabled or deleted provider, credential-unusable config, and invalid effort values are rejected with distinct tool errors before any child chat is created.
• type: "computer_use" rejects both args with a documented error; its model routing stays specialized.
• Spawns without the new args are byte-for-byte unchanged (personal override, deployment override, inheritance, explore current-turn fallback).
• Transcript: spawn rows with an explicit selection render Spawned {title} with {model display name}, {effort} thinking. Effort display mirrors backend resolution (requested wins over the config default, clamped to the config max, omitted for models without effort settings). Spawns without explicit args are unchanged. Resolution uses the current model config catalog, so later admin edits to a config can shift historic labels.
• Docs: docs/ai-coder/agents/{architecture,index,models}.md updated to describe spawn-time model selection.

No DB migrations, codersdk, or HTTP API changes; the selection reuses the existing LastModelConfigID persistence and enabled-config resolver.

Testing

• New internal tests in coderd/x/chatd/subagent_internal_test.go covering explicit selection, precedence over overrides, effort-only and combined spawns, all rejection paths, computer_use rejection, and list_subagent_models visibility filtering.
• Frontend: spawnModelDisplay.test.ts unit tests for the label resolver (default fallback, clamping, unknown config, none, out-of-scale values) plus Storybook play-function stories for the four label variants.
• go test ./coderd/x/chatd/ green; make lint, pnpm lint, and formatting clean.
• Manual UAT against a live develop.sh deployment: all six plan flows plus edge cases passed (children verified via DB last_model_config_id and initial-message reasoning effort).

This PR was written by Mux, an AI coding agent, acting on Mike's behalf.