fix: make Claude eval reruns scorable from Claude Code sessions

[fazxes]

What changed

• Added Claude Code session detection in nightshift test startup.
• When a nested Claude invocation would run inside Claude Code, the runner now falls back to codex if available, or fails early with an actionable message if it is not.
• Added a shared test-runtime-dir handoff so eval subprocesses and the parent eval runner read the same artifacts.
• Preserved the actual runtime agent in eval reports and added regressions for both the fallback path and the eval artifact handoff.
• Marked task #0277 done and recorded the fresh eval report.

Root cause

• The child nightshift test invocation could not reliably run inside the Claude Code shell when launched for eval reruns.
• The eval wrapper also needed a shared runtime directory so the parent could score the child run’s artifacts.

Validation

• make check
• Fresh eval rerun: .recursive/evaluations/0093.md

Result

• The fresh Phractal rerun now produces a scorable report instead of halting after two agent failures.
• The report records a fallback run using codex with a total score of 78/100.

[fazxes]

Closing this PR without merge. It failed the required code and safety review pair twice.

Blocking issues from the second review cycle:

• nightshift/infra/worktree.py: the NIGHTSHIFT_TEST_RUNTIME_DIR override still permits symlink/ownership redirection under the allowed temp prefix.
• nightshift/owl/eval_runner.py: the child eval subprocess still inherits a broad ambient environment.
• nightshift/tests/test_nightshift.py and nightshift/tests/test_eval_runner.py: the new tests still do not exercise the real fallback path end-to-end.

Per Brain protocol, task #0277 is being marked blocked after two failed fix-review cycles. A future attempt should start from these findings rather than reopening this PR.