[Birfy]

Closes #280

CladeThompson.tau() (#276/#279) mishandled three input edge cases:

• Non-finite budget_remaining: +inf (a caller's "no ceiling") and NaN returned max_tau, maximal exploitation from round one. They now read as a full budget (tau = 1), matching CostEfficient's clamp and BudgetGovernor.remaining_fraction()'s "no budget means 1.0". Only <= 0 (and -inf) means exhausted.
• Overflow: b ** -beta was evaluated before the cap, and a Python float ** raises OverflowError instead of returning inf (for example cool_down_beta=40, budget_remaining=1e-9). The exponent is now computed in log space and capped at max_tau before exponentiating.
• Constructor: NaN/inf cool_down_beta and NaN/inf max_tau slipped past the < 0 / < 1 checks; an infinite cap could hand random.betavariate an infinite tau. Both are rejected with ValueError now.

Existing test_tau_schedule expectations are unchanged. docs/api.md was regenerated and docs/selection.md documents the budget reading.

Tests: 3 new tests (7 cases) in tests/test_selection_hgm.py; 6 fail on main and all pass here. Full local suite passed.

[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 2026-10-15T08:38:06.775214Z	befeca9	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]

Review (Closes #280)

Verdict: LGTM — posting as COMMENT because GitHub blocks self-approve on Birfy→Birfy PRs.

Correctness

All three #280 failure modes are fixed and match the issue repro on main at HEAD:

1. Non-finite budget — +inf / NaN → tau = 1 (full exploration); only <= 0 / -inf → max_tau. Aligns with CostEfficient's clamp (max(0, min(1, budget)), which already maps +inf→1 and — via min(1.0, nan) argument order — NaN→1) and with BudgetGovernor.remaining_fraction() / default budget_remaining=1.0 when there is no ceiling.
2. Overflow — log-space log_tau = -β · log(b) capped before exp avoids the Python float ** OverflowError on steep β × tiny legal fractions (1e-9, even 5e-324). Schedule for in-range values unchanged (tau(0.002) ≈ 500 with β=1).
3. Constructor — not (isfinite(x) and …) correctly rejects NaN/±inf knobs that slipped past the old < 0 / < 1 checks; an infinite max_tau can no longer hand random.betavariate an infinite α/β.

Existing test_tau_schedule expectations are preserved; cool_down_beta == 0 short-circuit still returns 1 without touching the budget.

Design / docs fit

• Small, single-purpose fix on the surface added in #279; no API shape change.
• docs/selection.md documents the CostEfficient-aligned reading; docs/api.md gains the tau method row; CHANGELOG Fixed entry is clear and issue-linked.

Tests

Locally: tests/test_selection_hgm.py — 15 passed on this branch.
Against main code with the new test file: 6 failed, 9 passed (the three new tests / parametrize cases), matching the PR claim.

Coverage is tight: non-finite budgets, overflow + select() still works under a tiny budget, and constructor rejections including max_tau < 1.

Nits (non-blocking)

• Docstring says CostEfficient “reads” NaN as full; that is true for the current clamp expression order, but it is accidental (order-dependent min/max with NaN) rather than an explicit isnan branch. CladeThompson’s explicit math.isnan is the clearer contract — fine as written.
• docs/api.md table line for tau is one-line only; the non-finite convention lives in selection.md / the method docstring, which is enough.

CI

Checks were still pending / in progress at review time (build + test matrix 3.9–3.13). No merge from this review.

Ship it once CI is green.

[Birfy]

Re-review at c56fabc: still LGTM (COMMENT, since self-approve is blocked). This push only merges current main (3947aee, after #262 landed). The sole conflict was CHANGELOG.md, resolved by keeping both Fixed entries, so the branch diff against main is otherwise unchanged. tests/test_selection_hgm.py: 15 passed locally. Exact-head CI (build + 3.9–3.13) is green, the PR is MERGEABLE, the Codex review on befeca9 finished with no findings, and there is no open blocking feedback.

[Birfy]

Re-checked after merging main (924c25c, includes #272) at 42a9cd3. The only conflict was CHANGELOG.md (### Fixed): both entries are kept, the #280 CladeThompson.tau entry above the #271 runstore entry. No code conflicts. tests/test_selection_hgm.py and tests/test_selection.py pass locally on the merge commit, and CI is fully green (build + 3.9–3.13). No unresolved review threads. LGTM, still merge-ready.

[Birfy]

Merged main (751b8dd) into this branch — CHANGELOG-only conflict resolved (kept #280 + #286 Fixed entries). Head 669a196. tests/test_selection_hgm.py 15 passed locally.

LGTM once CI is green.