Winners-curse revs 3–4: the growth lookahead as a virtual injection + the yield rule

[gfrmin]

Code PR for the ratified winners-curse design (#195) plus its ratified-in-conversation §8 amendment (rev 4). Gate status: HALT THE LINE — opened for review, not for merge on this run (see the results commit and the findings below).

What this builds

• The union coherent injection (add_programs_to_state!(state, ::Vector{Grammar}, …)): one joint newcomers-only replay over the globally-deduped union of all candidates' programs — one ledger, one construction. Joint-not-sequential is load-bearing for exactness; commutation is inherited from Coherent injection: hypothesis addition commutes with conditioning (dominance halt-the-line fix) #187's theorem and asserted for the vector method (test_virtual_injection.jl §5). The single-grammar method's per-gid dedup is untouched (bit-stability of re-enumeration); the injection arithmetic is extracted verbatim into _inject_coherently!.
• GrowthProposal (scratch, n_added, yield_nats, p_newcomers) — one virtual injection, projected by the score, the op log, and the transition (Invariant 3). All candidates injected (Q3); the Occam charge rides inside the mixture (Q4 — GW_FEATURE_PRIOR_TERM deleted, no prior term at any seam).
• The yield rule (§8.2): score = net_value(injection_yield_nats(scratch), op_compute_cost) — the union-over-incumbent window Bayes factor under the complexity prior, net of the declared price. The rev-3 flow is never written: any own-posterior forward flow equals KL(P_union‖P_inc) ≥ 0 (Gibbs), so the §1b story was unsatisfiable by theorem — TDD caught it (§8.1). No ×H (the wait-option/Gittins argument — the far-term value is common to adopt-now and wait-and-re-decide), no plateau (realised yield already prices reality-of-gain).
• FIRE = adopt the scratch (adopt!, ratified Q1 memo via growth_cache): score ≡ transition as an === identity (T-3.55), pinned through the host seam. The cache epoch now bumps on every space change (the proposal depends on the live belief; a stale scratch must never be priced or adopted).
• T-3.53 exact in form: escape ops net_value(E[next yield], cost) (learned); growth ops net_value(computed yield, cost) (exact). One observable, one currency, one score form, two fidelities.
• Retired: _grammar_marginal_log_loss, the residual-mass screen, both per-candidate argmax cores, explore_grammar/explore_features/exploration_voi/exploration_fit/feature_discovery_voi/feature_discovery_fit, GW_FEATURE_PRIOR_TERM, the host's h_events/n_cond_events horizon bookkeeping (nothing consumed it — the §1 stationarity residual discharged by deletion at this seam).

Verification

• test_virtual_injection.jl (new, 27 checks): the seed-6 chance window waits at −0.56 nats while the retired score's counter-oracle fires at +5.12 on the same fixture; the informed fire lands from n = 8; multiplicity-by-mixture; commutation inheritance; Removal consumption: replacement semantics for applied removals (ratified #192) #193 self-healing end-to-end; score ≡ transition as ===.
• Re-baselines: test_threshold_explore §3, test_feature_discovery §3–§4 (stronger pin: realised yield strictly below the fit-axis Δℓ — the mixture pricing the prior axis), test_growth_returns §6, test_grid_world_meta §1/§3/§4. Full 56-file suite green; credence-lint 178 files, 0 violations.

Gate findings (the honest part)

The winner's curse itself is dead: every eu_max fire is informed (min yield 0.748, min P_newcomers 0.613 across verbose seeds — mechanism claim (i) verified), churn stays dead (13.1 metas/run), the headline vs never_explore holds (+0.106 CI > 0; final-regime 17–3, p = 0.0026), and the union replay is ~3× faster than the retired lookahead.

But 8 assertions still fail, and the must-not-degrade clause binds: the shared adopt-the-union transition made blind frequent firing much stronger for every policy (fixed_k5: 48.5 → 52.9 AUC, now significantly above eu_max on the primary). The ε-price counterfactual (20 seeds, reported-not-asserted) shows the cost slot is not the fix: eu_max recovers only +0.04, the escape tier churns (a learned E[yield] never reaches exactly 0 — only the price stops it), and the refinement-first gate starves features. The real finding is one level up: the gate's utility charges zero compute, so "move the boundary every 5 steps for free" dominates any priced selectivity by construction — the gate cannot reward the thing the mechanism optimises (eu_max: 13 metas/16.2s; fixed_k5: 42 metas/1.9s — invisible to the gate). Clairvoyant is now degenerate for the same reason (its forced at-change fires all no-op on the yield floor; trajectories identical to eu_max).

Proposed next move (needs ratification): a dominance-design amendment charging compute in the gate's utility — the one-currency exchange rate (evidence-nats price ↔ energy ↔ replay compute) declared as task data in advance — then re-run. This is measure-matches-declared-utility one level up: the measure set was fixed in advance, but the utility itself under-specifies compute.

🤖 Generated with Claude Code