Resolve managed coding-agent config over local state at launch

[asujithan]

Changes

Applies the admin-published managed coding-agent config at launch, taking precedence over the developer's own ucode settings.

On every ucode launch, the workspace's managed config is re-read and persisted to ~/.ucode/managed-state.json, then resolved against ~/.ucode/state.json per key — the manifest wins, the developer's state fills any gaps. The resolved view renders the agent settings file (e.g. ~/.claude/ucode-settings.json); the two state files are never merged on disk, so state.json keeps recording what the developer configured for themselves.

— managed_resolve.py — pure precedence logic. Claude resolves per model family (a family the admin didn't pin keeps the developer's choice); flat-list agents take the manifest's list wholesale.
— managed_config.py — adds the per-launch refresh. A failed read falls back to the last persisted config rather than blocking the launch; a successful "no config" read clears the persisted copy so a removed policy can't reappear after an outage.
— state.py — save_state swaps managed values back for the developer's own before writing, so precedence never overwrites state.json.
— cli.py — wires this into _launch_tool. An explicit --provider that conflicts with the admin's is a hard error; a managed provider that can't be used names the managed config in the error.

Followups:

• models and provider services, MCPs/skills, custom_headers, budget_policy/recommendModels, and sudo ownership of managed-state.json are follow-up PR
• ucode configure / first-run auto-configure still pull agent settings from local state in this PR; follow-up PR will gate configure when a managed config exists ("deprecated, run ucode" message + path)

Test plan

uv run pytest — 1173 passed, 6 skipped. ruff check . and ruff format --check . clean.
tests/test_managed_resolve.py
— TestClaudeModels (5) — proto slots → families; manifest wins per family; unpinned family keeps developer's value; falls back to local; None when neither side has models
— TestListModels (3) — manifest list replaces local; local stands when manifest silent; blank entries dropped
— TestManagedProviderService (3) — manifest-only read, no fallback to local state
— TestResolveState (4) — input state not mutated; managed values layered onto the copy; unrelated keys preserved; other tools' providers untouched
— TestStateFileIsNotRewritten (8) — the core contract, run against a real temp state.json with save_state unstubbed: developer's model survives; settings file still gets the managed model; overlay bookkeeping never on disk; repeated saves stay idempotent (the relayed-proxy path saves twice); provider_services preserved; a developer with no prior value isn't given one; codex + gemini parametrized

tests/test_managed_config.py
— TestRefreshManagedConfig (9) — persists and returns the manifest; no-config returns None; no-config clears the persisted copy; read failure and auth failure each fall back to persisted, or to local settings when there is none; empty persisted marker isn't treated as a fallback; no workspace is a no-op
— TestManagedLaunchState (2) — layers managed models when a config exists; hands state back untouched when none
— TestPersistence — empty config overwrites a previous one

Ran e2e tests locally against eng-ml-inference-us-east-1 workspace and all tests pass except tests/test_e2e.py::TestCodexLaunch::test_launch_codex_per_model which is expected due to codex model list fix not yet being rolled out to prod.