fix(capsule): clear stale maker patch before author reset

[Brian Krabach (bkrabach)]

Summary

Clear the stale .ai/hypothesis.patch scratch artifact during the capsule pipeline's author_reset, while preserving feedback and making cleanup failure loud. The change is one line in the shipped author_reset command plus three regression tests that extract and execute the shipped command rather than duplicating it. This is part of cross-repo dot-runner#79; it does not close that issue before the port ships. No engine-fidelity core or extension behavior changes, and no paid LLM semantic pipeline claim is made.

This is a toward-spec / graph-owned lifecycle hygiene fix: the surrounding attractor lifecycle describes reset/restart and context updates, while the graph owns this scratch-state cleanup. The change does not alter engine lifecycle semantics or imply an engine change. The relevant local context is the canonical lifecycle/reset description in specs/canonical/attractor-spec-canonical.md §3.1 and the repo's recorded graph-owned idempotency/lifecycle policy in SPEC_CONFORMANCE.md (ATX-2 and ATX-12). The reset remains loud: failure to prove the pinned clean state routes to reset_fail.

Verification checklist

• nlspec evidence: cited section(s) + holistic-context note (spec silence ≠ support; silence → extension process or design guidance)
• Unit tests pass (pytest modules/loop-pipeline/)
  • Applicable source-only invocation: uv run --isolated --with pyyaml --with pytest pytest -q tests/test_capsule_author_reset.py — 3 passed. The source-root prerequisite was supplied only for this test invocation; no production dependency was changed.
• Live pipeline run exercising changed code path — not applicable: this is a graph-owned shell reset command and regression fixture, with no engine code or handler change. Independent DTU validation extracted and executed the shipped command.
• AGENTS.md reviewed; repo-specific gates met
• Backward-compat path unchanged (if applicable)
• If this PR changes an observable contract — dispatch semantics, event contracts, admission/validation behavior, or anything else a pipeline author or downstream consumer could observe — it includes a specs/EXTENSIONS.md entry (new or updated), or this box is checked with an explicit one-line reason the change doesn't need one (e.g. "internal refactor, no observable change"). This changes only capsule scratch cleanup and adds no engine/extension contract.
• If this PR adds or changes a doc claim about code behavior (a number, default, vocabulary, or contract), it ships a guard test pinning that claim to its source of truth in code (see docs/OPERATIONS.md), or this box is checked with a one-line reason none is needed. No docs changed.
• Pre-publication leak review — no new public content class in this diff; fresh independent review found no public fixture leaks. Deterministic changed-file review found no local paths, credentials, or secrets.
• PR body includes verification evidence, not just "tests pass"
• CI is green before merge — on every path (auto-merge, manual/UI merge, or agentic/CLI merge). Confirm with gh pr checks <n> and look for CI Gate (all checks passed) reporting pass. This is checked after opening the PR; merge will not bypass red or pending CI.

Verification evidence

• RED control (pre-fix graph): CAPSULE_AUTHOR_RESET_DOT=<parent graph> python -m pytest tests/test_capsule_author_reset.py -q → 3 failed.
• GREEN control (c330 graph): python -m pytest tests/test_capsule_author_reset.py -q → 3 passed.
• Independent DTU validation: source test 3 passed, source root guards 253 passed, 2 skipped, source lint OK (no findings), Graphviz parse exit 0. The combined DTU validation also recorded port checks separately, but this PR publishes source only.
• The three fixtures cover regular stale patch deletion, dangling symlink deletion, and a directory cleanup failure that remains present and yields reset_unproven; feedback bytes are asserted unchanged in all cases.
• No raw logs or secrets are uploaded. Evidence source retained locally at .amplifier/evaluation/finish-round-20260908/dtu/validation-attempt3-independent.log.

Notes for reviewers

• Experience captured by this fix: repeated rounds must not reuse a previous patch; feedback is retained; cleanup failure is loud rather than silently absorbed.
• Scope is intentionally source-only. Do not publish the dot-runner port or make resolver changes as part of this PR.
• No paid LLM semantic pipeline run was used or claimed; the independent validation used deterministic tools and real container execution.
• Breaking changes: none.

Observations

None arose.

Part of dot-runner#79; do not close that issue until the port ships.