fix(worker): hard-stop reject-artifact patch failures (§60)

[SaJaToGu]

Closes §60 (docs/BACKLOG/open.md).

Bug

workers/execution.py classify_worker_outcome only hardened
against PARTIAL_PATCH_FAILURE_RETURN_CODE = 6 (PR #442 / §57).
It did not generalize to "any nonzero worker-returncode that
produces partial on-disk changes must be a hard stop".

returncode = 5 (reject artifacts in OpenRouterWorker, i.e.
.orig/.rej files left in the working tree after a partial
git apply) still fell through to the generic nonzero_with_changes
branch with failure_class: success. The run then proceeded to
commit + push + open a PR with whatever changes had been partially
applied.

Repro (PR #444)

Issue #390 validation run, 2026-06-26, gpt-4o via
openrouter_direct:

• worker_exit_code: 5 (Reject-Artefakte)
• run_outcome_worker_status: failed
• run_outcome_has_changes: True
• run_outcome_failure_class: success (lie classification)
• run_outcome_delivery_status: pr_created
• PR [AI] Fix: Periodic documentation benchmark with free OpenCode models *(parked)* #444 = 5-LOC trivial dummy function
• provider_scorecard_test_result: not_run (timeout 300s)

PR #444 was closed + branch ai/fix-issue-390 deleted. Same
symptom as PR #441 / §57, just one returncode class later.

Fix

• workers/base.py: new shared constant
  PATCH_VALIDATION_FAILED_RETURN_CODE = 5. Reason
  patch_validation_failed documented in the outcome taxonomy.
• workers/execution.py classify_worker_outcome: new
  branch between partial_patch_failure and changed —
  returncode == PATCH_VALIDATION_FAILED_RETURN_CODE →
  WorkerOutcome(False, has_changes, "patch_validation_failed").
  Returns before the generic nonzero_with_changes branch.
• workers/openrouter_worker.py: uses the shared constant
  instead of the magic 5 literal. Doc-string for the
  returncode semantics updated.
• scripts/solve_issues.py: doc-string for
  run_openrouter_direct_worker returncode semantics updated.

Scope

User-specified scope — targeted fix, no refactor:

In scope: Returncode 5 hard-stop.
Out of scope: the generic nonzero_with_changes semantics
for other workers. nonzero_with_changes may exist intentionally
to let some workers proceed for further review despite nonzero
exit. Do not refactor it without test evidence that it's
safe.

If other nonzero returncodes (e.g. 3 for timeout) need the same
treatment, that is a separate item (potential §60b).

Acceptance test

User-specified reproduction:

returncode=5
git_status=" M scripts/solve_issues.py"
→ WorkerOutcome(should_continue=False, has_changes=True, reason="patch_validation_failed")

Covered by tests/test_worker_execution.py (+1 new test) and
tests/test_worker_adapters.py (+1 new test).

Verification

• ./.venv/bin/python -m unittest tests.test_worker_execution -v: 22 OK (was 21, +1)
• ./.venv/bin/python -m unittest tests.test_worker_adapters -v: 96 OK (was 95, +1)
• ./.venv/bin/python -m unittest tests.test_openrouter_worker -v: 53 OK
• ./.venv/bin/python -m unittest tests.test_solve_issues -v: 168 OK (no regression)
• git diff --check develop..HEAD: clean

Diff: 6 files, +39/-5.