[Birfy]

Closes #302

What

• The default inner DefaultFusion is now always built with tournament=True; the outer LearnableFusionPolicy.tournament flag only selects the path (ranked vs. budget-degraded union path), so a False -> True flip (incl. the engine budget governor) reaches a real ranking.
• _run_tournament forces ranking on a caller-supplied inner left at tournament=False for that call only (restored after), and refuses to count a run / train a label if the inner still recorded only unranked trials.
• Budget-disabled union path (unranked trials, no labels) and the default behavior are unchanged.

Why: follow-up to the #298 review finding — after tournament=False -> True the inner returned an unranked union that was counted as a measured tournament and a positive label.

Tests (tests/test_learnable_fusion.py): False->True flip, True->False->True, epsilon=1 forced exploration, governor-style attribute flip, supplied unranked inner (restored after), inner that still returns unranked trials (no counters/labels). Instrumented verifier asserts real ranking calls; labels come from the measured winner. New tests fail on main; full pytest -q passes locally. Fragment: changelog.d/302-learnable-fusion-ranked-inner.md.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-16T11:55:45.157658Z	8e71220	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Birfy]

Automated review — #308 (Closes #302), head 8e71220

Verdict: LGTM ✅ (posted as a comment; self-approval is blocked, and auto-merge already squash-merged this at 20:10 CST, so the items below are follow-up nits, not blockers).

Correctness

• Root cause is addressed in both places it could bite: the default inner is now always DefaultFusion(tournament=True), and _run_tournament forces ranking on a caller-supplied inner left at tournament=False, restoring it in a finally. The outer tournament flag now only chooses the path (ranked vs _union_policy), which matches the governor's fp.tournament = should contract in evolution.py / async_evolve.py.
• I checked DefaultFusion.select: in tournament mode it always appends a ranked FusionTrial, and only _commit_union records ranked=False. So the new "no newly-recorded ranked trial → don't count, don't observe" guard never suppresses a legitimate label from the default inner. The existing default (tournament=True from construction) is byte-for-byte the same behaviour.
• The budget-disabled path is untouched: no labels, and unranked trials still land on the shared trials list that fusion_stats reads.

Tests (run locally, Python 3.13, fresh clone)

• tests/test_learnable_fusion.py passes on the branch. With agentdescent/fusion.py reverted to main, 5 of the 6 new tests fail, which confirms they reproduce #302. test_true_false_true_round_trip_ranks_on_the_second_enable passes on main too, because a True-built inner was never affected. That's fine as a regression guard, but "new tests fail on main" in the description is 5/6, not 6/6.
• Full pytest -q on the branch: everything passed except tests/test_era_algotune.py::test_the_reference_scores_about_one_through_the_sandbox, which then passed in isolation on both the branch and main. It looks like a load-sensitive sandbox timing flake, not related to this PR. CI is green on build + 3.9–3.13.

Nits / follow-ups

1. test_the_engine_governor_flip_to_on_ranks_through_evolve doesn't go through evolve. It reproduces the attribute write by hand, which is the same as the first test. #302 asked for the governor path specifically. A test that drives evolve (or async_evolve) with a budget that crosses the soft floor and then recovers would actually pin the integration. Otherwise, rename it so it doesn't overstate coverage.
2. Temporarily mutating a caller-supplied inner (inner.tournament = True … finally: False) is fine for the engine, because the fusion decision runs at the round barrier and the async comment says no worker reads the flag. Still, it isn't reentrant if a user shares one DefaultFusion instance elsewhere. A cleaner long-term shape is a ranked entry point on the inner (e.g. a _rank(artifact, diffs) helper that select calls when tournament=True), so the label path never has to flip shared state.
3. Unranked-only inners become silent no-op learners. With e.g. ReflectiveFusion as inner, which records ranked=False by contract and has no tournament attribute so it isn't forced, the wrapper now never counts a run or trains. That's correct per #302, but it would help to say so in the class docstring, or warn once in status(), so a user doesn't wonder why model_observations stays at 0.
4. The trials[before:] slice relies on trials being append-only. It's documented that way on DefaultFusion, but a short comment here noting the dependency would help future readers.
5. Docs: the tournament parameter's meaning changed slightly (it now picks the path, and no longer configures the inner). Consider one line in the class docstring / docs/api.md. The changelog fragment already describes it well.

Nice, tight fix: the scope stays within the acceptance criteria, and the fragment follows the new changelog.d convention.