test(phase3-C): end-to-end acceptance — unattended UI resolution cycle (#33)

[CameronCrow]

Resolves #33 — Phase 3 deliverable C, the end-to-end acceptance proof of the
unattended UI-resolution cycle, deferred until A (#35) and B (#32) landed (both now on main).

What this adds

tests/test_acceptance_phase3.py — the cohesive Phase-3 acceptance test. It drives the three
real cores wired exactly as Claude Code runs them, against a tmp_path project with synthetic
hook payloads (no live Claude Code runtime is needed to exercise the deterministic contract):

edit a UI file            -> PostToolUse flag_ui_change.flag() records it in pending-ui
agent tries to stop       -> Stop stop_gate.decide() BLOCKS (nothing verified yet)
cyclaudes verify runs      -> verify_result.write_result() records the outcome
agent tries to stop again -> Stop routes pass -> allow / fail -> block+diff / abstain -> allow+escalate

Scenarios (one per planning/PHASE_3.md success criterion)

• Happy unattended cycle (criterion 1): UI edit → flag → Stop blocks → pass covering the file → Stop allows. Zero human input.
• Self-correct loop (criterion 3): fail → Stop blocks with the expected-vs-actual diff → agent fixes → re-verify pass → allow. The break is caught, never reclassified as a pass.
• Non-UI change (criterion 4): a non-UI edit does not flag and the Stop gate does not block — the cheap relevance guard.
• Abstain escalates, never thrashes (criterion 2, the load-bearing rule): abstain makes Stop allow + escalate; 12 re-entries stay block_count == 0 and it is tallied once — proving it cannot consume the block budget toward Claude Code's 8-consecutive-block cap and false-pass there. A blocking abstain is the exact failure the phase exists to prevent.
• Fail bounded retry: blocks with the diff up to CYCLAUDES_RETRY_CAP, then escalates (allow + "after N attempts") rather than blocking forever.
• Wiring honesty: the exit-code → outcome mapping can't silently reclassify a real failure as pass/abstain, and the writer rejects an out-of-contract outcome.

Deterministic / fake-driven; green under the default python -m pytest (no desktop, no live runtime).

Test run

python -m pytest -q → 239 passed, 6 deselected (the 6 are -m live desktop checks).

Docs

Checked the Phase-3 success-criterion line in planning/TODO.md and noted C landed in planning/PLAN_MAIN.md. No other phases' TODO lines touched.

🤖 Generated with Claude Code