[linson007]

Measures proposal-policy alternatives against the same full target distribution and correct predecessor context. The reference compares existing, history-only and history-plus-prefix penalties on fixed candidate support, reporting overlap, target tail mass and proposal entropy. Prompt occurrences activate repetition; presence/frequency count generated output.

Related to #545. This implements the numerical measurement/reference stage; the research issue stays open for deployed distribution capture and serving qualification.

Validation:

• Full repository Python suite passed: 1325 tests, 27 expected skips, after building the required native binary in this worktree.
• Seven focused tests passed: target-tail preservation, signed penalties, default controls, within-block counts, greedy tie order, sampled frequencies, context mismatch rejection, and CLI/schema round-trip.
• The exact accept-plus-residual reference reproduces arbitrary full target distributions on 100 small cases, including disjoint proposal support and target tail tokens.
• Ruff lint/format and architecture boundaries passed; production native build succeeded locally.

python -m unittest dev.tests.tools.test_dflash_policy_overlap -v
python -m dev.benchmarks.dflash_policy_overlap capture.json --output report.json

Inputs must supply the actual full target p after request filtering and correctly bound proposal scores. The existing greedy capture in #550 does not contain that distribution. Fixture provenance is explicit and is not real-model improvement evidence. Target top-k/top-p/min-p, stops, thinking and constraints are already reflected in p; this tool changes no runtime policy, retrieval support, proposal q buffer or rejection correction. The CPU reference is not FP32/Metal numerical parity. Native sampled capture, mixed-lane/state tests and paired latency/acceptance measurements remain open.

[linson007]

Validation scope relative to the new RunPod/native test run

This PR remains a CPU numerical proposal-policy reference, not a runtime policy change. Its recorded qualification includes seven focused policy tests, full-target/tail preservation and 100 small accept-plus-residual distribution cases, plus the passing repository suite and current-head scheduled CI at 24b00c1f5b22d0ae4378980c273a3478227bee50.

The new real-Qwen3.8 matrix in #554 used greedy decoding, all verification rows and frozen budget learning. It did not exercise sampled proposal-policy alternatives and does not supply the actual full filtered target distribution this reference requires. Consequently no new GPU/model-level policy-effectiveness result is claimed for this PR. Sampled native capture and paired latency/acceptance qualification remain open in #545.

Full report and reproduction