[Birfy]

Closes #872

The plan / algorithms run preflight for a reflect: openai_compatible block with an unset key always said the run "fails after two rounds". The loop actually stops after min(evolve.rounds, PROPOSE_FAIL_ROUNDS) rounds, so the --rounds 1 --workers 1 quick trial was told the wrong thing.

• _propose_fail_rounds_text(spec) reads PROPOSE_FAIL_ROUNDS from evolution and the spec's evolve.rounds; wording is unchanged when rounds is unset or ≥2 (existing exact-string test still passes).
• New parametrised test for rounds 1/2/8.
• changelog.d/872-dead-reflector-rounds.md.

Local: pytest -k "cli or preflight or plan or algorithms" green (with ARK_API_KEY unset; test_present_host_passes is env-sensitive to that key on this machine, unrelated).

[Birfy]

Verdict: PASS (tip f74c6e0)

Reviewed the diff: the helper only changes the wording of a warning, clamps to PROPOSE_FAIL_ROUNDS from evolution (no duplicate constant), ignores bools/non-int rounds, and leaves the default text byte-identical (existing exact-string test). New test covers 1/2/8 rounds. risk_gate: T2. No behaviour change in the evolve loop.