feat(providers): add Requesty, MiniMax and Forge

[Zongwei9888]

Description

Ports three provider contributions that could not merge as submitted. #138 and #130 also touched new_ui/ and nanobot/; #116 targeted nanobot/ only. Those trees were removed in the v2.0 refactor — but the provider work itself applies unchanged to core/providers.

Supersedes #138, #130 and #116.

Why a dedicated ProviderSpec, rather than telling users to use custom

Each spec unlocks behaviour custom cannot express:

Provider	Spec field	What it buys
Forge (#116)	strip_model_prefix=True	Forge resolves bare model ids. Configured through custom, a habitual anthropic/claude-3 would be sent verbatim and fail.
Requesty (#138)	supports_prompt_caching=True	Injects cache_control for Claude models routed through it. Through custom there is no caching — a real cost difference for an agent replaying long contexts.
MiniMax (#130)	keywords, default_api_base	Convenience: MiniMax-M2 selects the provider automatically and the base URL is preset. Lowest marginal value of the three; custom works.

Both halves, not just the registry

config.py resolves providers with getattr(self.providers, spec.name, None), so a ProviderSpec without a matching ProvidersConfig field is silently skipped everywhere. All three get the registry entry and the config field, as the ProvidersConfig docstring requires. zhipu is the reference for the minimal-complete shape.

Endpoints verified reachable: router.requesty.ai (200), api.minimax.io (401), api.forge.tensorblock.co (401).

Bonus: a gateway bug #116 found, which core had half-fixed

_setup_env() forced spec.env_key for gateways but still used setdefault for env_extras:

if spec.is_gateway:
    os.environ[spec.env_key] = api_key      # gateway wins
else:
    os.environ.setdefault(spec.env_key, api_key)
...
for env_name, env_val in spec.env_extras:
    os.environ.setdefault(env_name, resolved)   # gateway does NOT win

A stale ambient variable therefore outranked the gateway the user had selected. #116 fixed exactly this in nanobot's provider; the same rule now applies to env_extras here. Latent on main today (only zhipu uses env_extras, and it is not a gateway), but it would bite the first gateway that needs one.

What was dropped from the originals

The tests covering new_ui's requesty_models service and nanobot's registry. One of them did pytest.importorskip("nanobot.config.schema"), which now resolves to an unrelated PyPI package (nanobot-ai) when that happens to be installed — so it fails locally and skips in CI, testing nothing either way.

Testing

• Full suite: 1154 passed, 3 skipped (main baseline 1131; +23 = Requesty 5, MiniMax 12, Forge 6)
• pre-commit run --from-ref origin/main --to-ref HEAD — clean
• The env_extras fix is verified in reverse: reverting it fails test_gateway_env_extras_override_ambient_values

Not verified: strip_model_prefix=True for Forge. Confirming it needs a Forge API key. The value is carried over from the author's nanobot spec and matches how AiHubMix was handled in that registry. It only affects users who explicitly configure Forge.

Unrelated flake spotted while validating

tests/application/test_automation_goal_runs.py::test_legacy_unreserved_turn_is_never_adopted_as_automation_initial_turn fails intermittently with timed out waiting for Run ... to become completed. Measured in isolation: 2/20 on clean main, 0/20 on this branch. Pre-existing, not from this PR — but it will redden CI at random and deserves its own issue.