fix(workflows): fail a gate whose on_reject is not abort/skip/retry instead of silently completing a rejection

[jawwad-ali]

Problem

GateStep.execute reads on_reject = config.get("on_reject", "abort"). In the reject branch it handles only two literals before falling through:

if on_reject == "abort":  ...
if on_reject == "retry":  ...
# on_reject == "skip" → completed, downstream steps decide

So any other value lands in the "skip" case. A gate that was rejected reports COMPLETED, and the run continues past the review the gate exists to enforce.

Reproduction on current main (81bf741)

Same config, REJECT verdict, varying only on_reject:

on_reject='abort'  -> failed     error="Gate rejected by user at step 'g'"
on_reject='retry'  -> paused
on_reject='skip'   -> completed  (by design)
on_reject='Abort'  -> completed  <-- rejection silently discarded
on_reject='fail'   -> completed  <-- same
on_reject='stop'   -> completed  <-- same
on_reject=None     -> completed  <-- same
on_reject=5        -> completed  <-- same

This is a fail-open failure mode on a safety control, and every trigger is an ordinary authoring mistake:

• a capitalisation slip — Abort, SKIP
• a guessed verb — fail, stop
• a bare on_reject: in YAML, which parses to None. Note config.get(key, default) does not substitute the default for an explicit null.

Fix

validate already rejects anything outside the three literals:

on_reject = config.get("on_reject", "abort")
if on_reject not in ("abort", "skip", "retry"):
    errors.append(...)

…but the engine does not auto-validate before execute (see WorkflowEngine.load_workflow), so this PR adds the matching execute-side guard. It mirrors the options and verdict_input guards already in this method, and is placed before the non-TTY short-circuit so the error surfaces in CI rather than pausing and failing later on interactive resume.

No breaking change. All three documented values behave exactly as before — the existing test_reject_verdict_input_preserves_reject_behavior parametrizes over them and still passes untouched. The only inputs whose behaviour changes are ones that previously discarded a rejection.

Verification

• Fail-before / pass-after: the seven parametrized invalid values fail on unpatched src and pass with the fix. tests/test_workflows.py: 27 failed → 20 failed, 837 → 844 passed.
• The remaining 20 are identical to the clean-main baseline I captured on 81bf741 (all Windows symlink-privilege).
• uvx ruff@0.15.0 check src tests → clean

Test goes into the existing TestGateStep, beside test_validate_invalid_on_reject — the validate-side half of the same contract.

---

Written with assistance from Claude Code. Bug found, reproduced, and verified by me on current main.