Fix: cancellation wiring

[tempusfrangit]

Summary

Wire cancellation end-to-end from HTTP to worker subprocess, and make CancelationException a proper public API.

Cancel was accepted by the HTTP layer but never reached the worker. This wires the full path: HTTP cancel/connection-drop → supervisor → orchestrator → ControlRequest::Cancel → worker → PyThreadState_SetAsyncExc.

Key changes

Cancellation wiring (existing commits)

• Orchestrator: add cancel_by_prediction_id (resolves pred ID → slot ID)
• Supervisor: cancel() delegates to orchestrator via spawned task
• SyncPredictionGuard: calls supervisor.cancel() on drop (not just token)
• Sync HTTP handler: spawns prediction in background task so slot lifetime is not tied to the HTTP connection (fixes permit leak on disconnect)
• worker_bridge: capture py_thread_id at prediction start, use it for cancel_sync_thread(); async cancel still uses future.cancel()
• Fix sync cancel status: upgrade PyErr to Cancelled when slot is marked cancelled

CancelationException as BaseException + public SDK export

• CancelationException now derives from BaseException (not Exception) so bare except Exception blocks cannot swallow cancellation — matching KeyboardInterrupt and asyncio.CancelledError
• Defined as a static type via pyo3_stub_gen::create_exception! (replaces dynamic builtins.type() approach)
• Exported via coglet._impl → coglet → cog.exceptions → cog.CancelationException
• Auto-generated type stubs include proper class CancelationException(BaseException) definition

Dead code removal

• Removed SIGUSR1 signal handler, CANCELABLE flag, CancelableGuard, enter_cancelable() — all dead since PyThreadState_SetAsyncExc replaced SIGUSR1
• Removed _is_cancelable Python getter from Server
• Updated is_cancelation_exception to use the static type instead of dynamic import from nonexistent cog.server.exceptions

Documentation

• New Cancellation section in docs/python.md documenting CancelationException usage, import paths, BaseException semantics, and re-raise requirement
• Updated docs/http.md cancel endpoint docs to use cog.CancelationException import path
• Regenerated docs/llms.txt

Tests

• 5 new Python cancel tests (explicit cancel sync/async, connection-drop)
• 2 txtar integration tests

closes: #2748

[Copilot]

Pull request overview

This PR implements end-to-end cancellation wiring from the HTTP layer to the worker subprocess, addressing issue #2748. The implementation replaces the previous SIGUSR1 signal-based approach with PyThreadState_SetAsyncExc for sync predictions, and makes CancelationException a proper public SDK export as a BaseException subclass.

Changes:

• Wires cancellation through supervisor → orchestrator → worker using cancel_by_prediction_id() and ControlRequest::Cancel
• Implements CancelationException as a static BaseException type (not Exception) exported through the public SDK
• Refactors sync HTTP prediction handling to spawn work in background tasks, using SyncPredictionGuard for connection-drop cancellation
• Removes dead SIGUSR1 signal handler code, CANCELABLE flag, and CancelableGuard
• Adds comprehensive documentation and integration/unit tests

Reviewed changes

Copilot reviewed 20 out of 21 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
python/cog/exceptions.py	New public exceptions module exporting CancelationException
python/cog/__init__.py	Adds CancelationException to public API exports
integration-tests/tests/cancel_sync_prediction.txtar	Integration test for sync prediction cancellation via HTTP endpoint
integration-tests/tests/cancel_async_prediction.txtar	Integration test for async prediction cancellation via HTTP endpoint
docs/python.md	New Cancellation section documenting CancelationException usage and semantics
docs/llms.txt	Regenerated documentation with cancellation updates
docs/http.md	Updated cancel endpoint documentation with correct import paths
crates/coglet/src/transport/http/routes.rs	Refactors sync predictions to spawn in background tasks with SyncPredictionGuard
crates/coglet/src/supervisor.rs	Implements cancel() method that delegates to orchestrator
crates/coglet/src/service.rs	Sets orchestrator on supervisor for cancel delegation
crates/coglet/src/orchestrator.rs	Adds cancel_by_prediction_id() trait method and event loop cancel handler
crates/coglet-python/tests/test_coglet.py	Adds Python unit tests for sync/async cancellation and connection drops
crates/coglet-python/src/worker_bridge.rs	Captures py_thread_id at prediction start, implements cancel via PyThreadState_SetAsyncExc
crates/coglet-python/src/predictor.rs	Updates is_cancelation_exception to use static type
crates/coglet-python/src/lib.rs	Exports CancelationException and removes dead signal handler code
crates/coglet-python/src/cancel.rs	Complete rewrite: implements PyThreadState_SetAsyncExc-based cancellation, removes SIGUSR1
crates/coglet-python/src/bin/stub_gen.rs	Adds CancelationException to public re-exports
crates/coglet-python/coglet/_impl.pyi	Generated stub with CancelationException class definition
crates/coglet-python/coglet/__init__.pyi	Generated stub re-exporting CancelationException
crates/coglet-python/coglet/__init__.py	Exports CancelationException from _impl
.github/workflows/ci.yaml	Adds coglet wheel download to lint-python job with proper dependencies

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The generated stub file is missing the header comments that stub_gen.rs claims to add. According to stub_gen.rs lines 54-56, the generated file should start with:

# This file is automatically generated by stub_gen
# ruff: noqa: E501, F401

However, the actual file doesn't have these comments. This suggests either the stub generation wasn't run after the changes, or there's an issue with the generation logic. This file should be regenerated to match the stub_gen configuration.