Runner treats porcelain writes as reads and varies classification by plan shape #60

On main d3d3b35, tools/agent_run.py::build_index omits porcelain operations from writer classification. An argv-only plan uses the registry; adding a use step switches to the catalogue. Neither path treats session-end --commit-message fixture --push as a writer.

In synthetic fixture regressions, the production runner adapter dispatched mocked session-end commit/push after a simulated lost gateway response under on_error: continue. The observed sequence was seal, session-end, inspect; expected was seal, inspect. Sessionless argv, prefixed argv and operation-id forms were admitted. No actual commit/push or live learner mutation was performed. Environment: Python 3.14, macOS, isolated checkout.

Repair on codex/runner-write-classification-20261027: use one executing parser/registry rule for both paths. Keep established query/tool reads; treat commands, porcelain, absent registrations and unknown kinds as write barriers. Preserve sessionless descriptions.

Acceptance:

• Writer sets match across task-resolution paths, including missing and unknown registry kinds.
• Refuse sessionless effectful steps before any dispatch, including mixed plans, prefixed argv and write: false.
• Block porcelain after an uncertain write while allowing independent reads.
• Descriptions of effectful operations remain sessionless reads.

Reproduce the unsafe behavior by applying the new tests to the base runner:

.venv/bin/python -m pytest -q tests/test_agent_run.py -k 'classification_is_conservative or session_end_commit_push or uncertain_write_blocks_porcelain or describing_porcelain'

Thirteen initial regression cases failed against the base, one passed. Dispatch and the lost subprocess response are mocked; the runner admission/scheduler and plan compilation use the production route, without an injected executor. The current focused runner/read-session/readiness/consolidation/task suites pass: 127 tests. Hosted CI will be recorded on the draft PR.

Retained evidence: work/complaints/2026-10-27-agent-run-write-classification.md.