dream(darwin-evolution): ADR-249's cost seam wired into evolve() (evaluated)

[ruvnet]

MetaHarness Dream Cycle, 2026-08-17. Deep surface: darwin-evolution (slot 2). Full detail, receipts, candidate-selection reasoning, adversarial-critique writeup, and weight-eft/learn scan findings: #206.

Hypothesis

Given evolve.ts's evaluateVariant, which already computes variantBytes(variant.dir) (a deterministic surface-size signal) elsewhere in the file for Pareto tie-breaking but feeds no signal into scoreVariant's ADR-249 signals.cost seam, when EvolutionConfig gains an optional costBudgetBytes field and, only when it is set, evaluateVariant passes { cost: { units: variantBytes(variant.dir), budgetUnits: costBudgetBytes } } into scoreVariant, then a variant whose surface files grow past costBudgetBytes should score a strictly lower costEfficiency (and, on a crafted near-tie fixture, this alone should be able to flip a promotion decision from promoted to not-promoted), while every existing evolve() caller — none of which set costBudgetBytes — continues to produce byte-identical ScoreCards to before, subject to: (a) zero regression in the existing 632-test suite, (b) omitted-field behavior stays byte-identical (ADR-249's own zero-cost-adoption contract, verified one level up), (c) scorer.ts itself untouched, (d) diff materially under 300 lines.

Candidate

ADR-249 (2026-08-10) shipped an opt-in signals.cost = {units, budgetUnits} seam on the frozen scorer, tested in isolation (scorer-signals.test.ts), but evolve.ts's only production call site never passed signals — the seam was structurally unreachable from any real evolution run. Fixed: EvolutionConfig.costBudgetBytes (optional), threaded through evaluateVariant into the existing scoreVariant call using the already-computed variantBytes(variant.dir) parsimony signal (previously used only for 'pareto' selection's tie-break). evaluateVariant is now exported (was module-private) for direct testability. 2 source files (evolve.ts +18/-2, types.ts +11), 1 test file (+79 lines, 4 new tests). Well under the 300-line target.

Evaluation Receipt

Deterministic reproduction (not sampled):

• Before/after repro: with the source fix stashed (test file kept), the 4 new tests fail exactly as predicted (evaluateVariant is not a function); the 5 pre-existing evolve.test.ts tests are unaffected.
• Fix restored: packages/darwin-mode full suite 632 passed, 14 skipped, 0 regressions. tsc --noEmit clean. Full monorepo npm run build clean (no wasm/NAPI degradation tonight).
• Two dependent packages (create-agent-harness, flywheel) that reference @metaharness/darwin types re-run clean (11/11).
• New tests assert the arithmetic directly: a 200-byte variant vs. a 100-byte budget decays costEfficiency to exactly round6(100/200) = 0.5; on a crafted near-tie fixture (parent finalScore=0.4, promotionDelta=0.02) that decay alone moves finalScore from 0.435 (promoted) to 0.385 (not promoted) — no other gate clause moves.

Baseline Comparison

No caller in the shipped orchestration sets costBudgetBytes, so every existing evolve() invocation is byte-identical to before this diff — verified directly by tracing the omitted-field path (not just via the pre-existing scorer-signals.test.ts fixtures, which only cover scoreVariant in isolation) and confirmed independently by the adversarial critic.

Darwin Lineage

Not run — this is a scorer-wiring/config-plumbing fix to the evolution loop's own orchestration code, not a tunable routing/topology/prompt/tool/tier parameter; no Darwin mutation surface applies (ADR-071 mutation allowlist), same reasoning as the 2026-08-15/16 entries.

Flywheel Evidence

Not applicable in the "ran a flywheel evolution" sense — this candidate modifies @metaharness/darwin's own scorer-seam wiring. Durable evidence is the deterministic test repro above, committed to docs/dream-cycle/2026-08-17-gist.md.

Reward Hack Check

Purely additive diff, no existing test weakened, no gold/benchmark data touched. variantBytes is the same signal already used for Pareto tie-breaking before this diff — if it's gameable by whitespace-stripping mutations, that property predates tonight. This diff does make it more consequential (now inside the primary score, not just a secondary tie-break) — disclosed, not hidden (see Adversarial Critique in the gist).

Security Review

No credential/shell/network surface touched, no new dependency. Also statically audited tonight (unrelated to this candidate, $0): darwin-mode's SWE-bench solvers (bench/swebench/*.mjs) all git fetch --depth 1 the base commit only, falling back to --depth 200 ancestor history on failure — no future/gold-patch commits are ever fetched, unlike the git-history-leak class of grading defect that hit SWE-bench Pro's public graders. Checked and clear, not a finding requiring a fix.

Regression Analysis

Zero regression: 632/632 packages/darwin-mode tests pass (0 failures, 14 pre-existing skips unaffected), tsc and full repo build clean, 2 dependent packages' darwin-referencing tests unaffected (11/11).

ADR

Appended a dated 2026-08-17 section to ADR-249 (the ADR that shipped the seam this PR wires up) rather than minting a new ADR number, matching the 2026-08-13/14 precedent for a coverage/completion addition to an existing architectural decision.

Research Gist

No gist-creation tool is available in this session's toolset (matching 2026-08-13 through 08-16) — committed instead: docs/dream-cycle/2026-08-17-gist.md (GIST=LOCAL).

Issue

#206 (full findings, candidate-selection reasoning, weight-eft/learn scan results, competitor review, ledger check).

Witness

session_commit : 5453c8c990824b54e05f289774e5a8b2cea0a32e
report_sha256  : b5e71a1f847cc37d43cb4c7b41db718d0832cec67aab701a121a0c3527b69693
witness        : 57ed9167c8085b57799a0479d9842baed13f94e782c889b9f6f00f0804f25a72

Verifier procedure is in the gist's Witness section and was independently self-verified (round-trip reproduced programmatically) before this PR was opened.

Merge Policy

Human review required. Do not self-merge. Do not autonomously promote Flywheel state.

Recommendation: ACCEPT-WITH-CAVEATS — real, independently-critiqued, fully deterministic and reproducible fix that closes a genuine "ADR shipped, implementation dangling" gap. Two comparable-severity gaps are disclosed as still-open, natural next-night follow-ups (not silently fixed under review pressure, not silently left undisclosed): (1) no real evolve() caller sets costBudgetBytes yet — this PR is plumbing only; (2) variantBytes's pre-existing gameability is now score-consequential rather than merely tie-break-consequential once a caller does opt in.

---

Generated by Claude Code