Module selector 

Token-budget selection over ranked context candidates (Wave 14B/15.1).

Four deterministic, no-panic selectors:

• select_with_budget: greedy value/token knapsack over ContextItems with a soft budget + hard maximum; required items are always kept (the caller compresses them when they alone exceed the hard max).
• select_in_order: the optimizer-off variant — rank-order cut, no value/token reordering.
• mmr_diversify: Maximal Marginal Relevance — penalizes candidates similar to already-selected ones so one component/owner group cannot crowd out the rest.
• enforce_quotas: TOKEN-aware per-kind caps (fraction × available tokens, adapted to the pool’s running average token cost) over a ranked list, preserving rank order.

Functions

enforce_quotas

Enforce per-kind quotas over a ranked list, preserving rank order — TOKEN-aware (reviewer item 6): each kind’s cap is a token budget fraction * available_tokens; a candidate is accepted only while accepting it would not push its kind’s accumulated spend over that cap. Because the cap is a token budget applied over actual per- candidate costs, the effective entry allowance adapts to the selected pool’s running average token cost (expensive candidates exhaust a kind’s allocation faster, cheap ones admit more entries). A candidate that would exceed its group cap is skipped — the next candidate of another group takes its place, so the remaining groups rebalance.

mmr_diversify

Maximal Marginal Relevance selection: repeatedly pick the candidate maximizing lambda * score - (1 - lambda) * max_similarity(selected), up to budget items. lambda is clamped to [0, 1]; 0.5 is the default (see mmr_diversify_default). Deterministic: ties keep the earlier candidate in ranked order.

mmr_diversify_default

mmr_diversify with the spec’s default lambda = 0.5.

select_in_order

Budget selection in list order (the optimizer-off ablation of select_with_budget): walk items in importance order, keep required items unconditionally, and accept pool items while the total token spend stays within both the soft budget and the hard_max ceiling. No value/token reordering — a candidate’s rank order decides.

select_with_budget

Select indices into items within a token budget. Required items come first and are never dropped (even when they alone exceed the budget — required context is never silently cut); the remaining budget is filled greedily by value/token, highest first. Deterministic: ties break on index order.