[anth-volk]

Fixes #57

Part of #53.

What this does

The first of the pull requests stacked on #55 that fix the Ruff findings #55 leaves ignored: the 469 findings in EM, RSE and TRY, after which pyproject.toml no longer ignores those three families. Behaviour is unchanged.

• EM101, EM102 (240): Ruff's fix assigns each exception's argument to a variable before the raise. The variable is named for what it holds, not Ruff's msg: startup_reason for the 19 startup reasons passed to EmitterStartupError and ServiceStartupError, refusal_code for the 38 refusal codes passed to RequestRefusedError and _refuse, and error_message for the other 181 messages. This also clears all 171 TRY003 findings, which flag a long message written in the raise.
• RSE102 (2): no parentheses on a bare raise TimeoutError (Ruff's fix).
• TRY301 (43): checks that raised inside a try block move into helper functions that the block calls, so each handler still catches the same exceptions in the same order:
  • the host's socket client (_encode_request, _check_response_size, _decode_response) and the interface's transport (the same three, the last named _complete_response);
  • the host runtime: _respond and _handle_request keep their catch-all handlers and call _answer and _run_request, with _read_request and _encode_reply as module functions;
  • the host entry point (_read_configurations, _load_modules);
  • both launchers (_configuration_payload, _send_configuration, _wait_for_readiness). The startup phase is still set before each step, so failures are classified as before;
  • the Orrery delivery (_check_unchanged, _upload_list, _check_listed_name, _check_upload_url, _check_publication_answer). Each helper takes only the values it checks, typed as the unchecked JSON values they are (object, or the SplitResult of the upload URL). The .get and urlsplit calls stay at the call sites, so every exception is raised where it was, and each upload's URL is still checked just before that upload;
  • the Hugging Face request parser (_file_entries), and one startup test, which raises from a local function.
• TRY300 (2): the host entry point and the Orrery delivery return success from an else block.
• TRY004 (11): a type check that raises something other than TypeError.

  • Four now raise TypeError, where no caller can tell the difference:

    • the runtime's request reader, whose caller answers every failure with the same error reply;
    • the entry point's module list, whose caller prints the same fixed startup reason for every failure;
    • the Orrery upload list, whose handler records ValueError and TypeError with one error code;
    • the telemetry block request. In the emitter it reaches the runtime's catch-all. It also reaches Microcosm's test support, whose fake telemetry emitter calls process_request directly, but there the branch can't be reached: the emitter's client always sends blocking_gate_ids as a list.

    One test now pins each caller's visible outcome (see the round 1 record below); the runtime's was already tested.

  • Seven keep their type, with # noqa: TRY004 and the reason in a comment above:

    • three in the interface's request objects, whose published constructors raise ValueError for every invalid value (two are tested with pytest.raises(ValueError));
    • the host client, which reports the exception's type name;
    • the JSON column type, which raises the same type as json.loads's JSONDecodeError for invalid JSON just above it;
    • the two Alembic environments, where the check is for a connection the emitter didn't supply.

    The last three were optional. Each could become TypeError without a suppression, but that would change the type a caller sees, which this pull request avoids.

Test-first record

There is no red step, since nothing changes behaviour.

Commit	Result
abc82ff The fixes	ruff check . and ruff format --check . pass with EM, RSE and TRY selected. All five guards pass with GITHUB_BASE_REF=main. The 25 test functions this commit changes (52 cases, by node ID) pass.

Review round 1:

Commit	Result
e2d8cb6 Merge #55's round 2	ruff check ., ruff format --check . and the guards pass.
33b01b8 Tests of the three callers	No red step, since they pin behaviour that doesn't change. They pass on this branch and, in a worktree of #55's head (bef4f36, where the three checks raised ValueError), with only the three test files copied in: 18 cases each time.
fa38ac1 Names, Orrery types, reasons	ruff check ., ruff format --check . and the guards pass. The four changed test functions pass (19 cases).

No other test was run: the source refactors are left to CI's test jobs. The node IDs exclude tests/test_collector_contract.py::test_lost_ack_retries_same_event_after_restart_without_duplicates, which needs the pinned collector.

CI

No workflow, trigger or script changes, and no environment variable or secret is read differently.

Release

The fragment is changelog.d/ruff-exceptions.changed.md, a patch bump: merging releases the next patch version of both distributions, whose code changes without changing behaviour.

🤖 Generated with Claude Code