fix: stop the README capture snippet from hiding failed builds

[PabloCodes7]

The bug

The README's flagship snippet is what a maintainer copies into their workflow:

- name: Run tests
  run: pytest -q | tee test.log     # <- exits 0 even when pytest fails

GitHub's default run: shell is bash -e — no pipefail — so a pipeline reports the status of its last command (tee). A failing test suite exits 0: the step goes green, if: failure() never fires, the action never runs, and the user's red build is reported as passing. Verified: bash -e -c 'false | tee x' → exit 0; bash -eo pipefail -c 'false | tee x' → exit 1.

The missing 2>&1 is the second half. Tools that report only on stderr (compilers, linters, mypy, cargo) leave the log empty, patchrail ci explain exits 2 without writing --out, and set -euo pipefail in action.yml killed the triage step with Process completed with exit code 2 — a second, meaningless red step stacked on the failure the user actually came to debug. A log-path pointing at a file that was never written was worse still: it fell through to ::error::provide either log-path or log-text, which is wrong — they did provide it.

The fix

• README: both snippets capture with shell: bash (bash -eo pipefail) and 2>&1, plus a short section explaining why each half matters.
• action.yml: a missing or empty log is a capture mistake, not a build failure. It now emits a warning annotation naming the cause and the correct capture recipe, sets empty outputs (guide-url = guide index), and leaves the step green. A workflow that configures no log at all still fails, as it should — the action can never work.

Guards

• tests/test_readme_snippets.py — parses every YAML snippet in the README and fails if a capture step drops pipefail or 2>&1, or tees a file it never hands to the action. Mutation-tested: each of the three regressions turns exactly one test red.
• tests/test_no_classification.py — the no-log path: exit 0, one-line annotation, empty outputs, summary explains the capture.
• smoke-no-log CI job — runs the action on a real runner against an empty log and a never-written log, asserts neither fails the job. It's the shell in action.yml, not the Python, that decides whether the step survives, so this is checked end-to-end. Added to sync-v1's needs, so the v1 tag only moves when it passes.

Happy path is unchanged (verified locally: same class, confidence and guide URL).