feat(gates): neutralizationGate — footprint-matched placebo control (0.107.0)

[drewstone]

What

Adds the one primitive the substrate genuinely lacks: a placebo control that proves a held-out win comes from the candidate's content, not from the prompt/mount footprint the content happened to add.

A held-out gate (heldOutGate/defaultProductionGate) proves candidate beat baseline. It cannot tell apart (a) informative content from (b) simply more bytes/lines the model spends attention on. neutralizationGate closes that hole.

How

• neutralizeText (src/campaign/neutralize.ts) — blanks every non-whitespace char to a 1-byte filler while preserving layout/length. Byte-exact for ASCII; character/layout-preserving for multibyte (what the tokenizer footprint tracks).
• neutralizationGate (src/campaign/gates/neutralization-gate.ts) — composable Gate. Compares the candidate's held-out lift against a footprint-matched neutralized variant's lift (reusing pairHoldout). Holds when the neutralized variant reproduces ≥ maxDecorativeFraction (default 0.5) of the lift — the "R46" decorative-reject rule — however large the raw lift. Fail-closed on no positive lift; fail-loud if composed without the loop wiring.
• GateContext gains optional neutralizedJudgeScores / neutralizedArtifacts.
• runImprovementLoop gains an optional neutralize fn; when supplied and the winner changed, it scores a third holdout arm and exposes it to the gate. Opt-in — one extra holdout campaign only when wired.

Compose after significance: composeGate(heldOutGate({...}), neutralizationGate({...})) — significance says the lift is real; this says the content caused it.

Provenance

Lifted from a hand-rolled tax-benchmark fork (compile-and-prove/ablation-gate). A three-agent audit confirmed the fork's other halves — its permutation lift test, <12 power check, JSONL ledger, content-hash promotion, and mine→gate→promote loop — are duplicates of existing substrate primitives (heldOutGate, powerPreflight, LoopProvenanceRecord, surfaceContentHash, selfImprove) and dissolve into them. This placebo control was the only piece with no substrate equivalent (grep-confirmed: agent-eval's "decorative" uses are dead-code detection + rubric validity, not a footprint-matched content control).

Verify

• pnpm typecheck → 0 errors; pnpm test → 272 files, 2797 passed; pnpm build → clean.
• New gate suite: 8 tests (ships-on-causal, holds-on-decorative, threshold equality, fail-closed no-lift, fail-loud missing-arm, neutralizer layout/byte behavior). Full gate suite 65/65; loop/distillation 37/38 (1 pre-existing skip) — wiring didn't regress the loop.
• Additive only (new exports + optional fields) → minor bump 0.106.3 → 0.107.0, no consumer lockstep. Version trio bumped.

[tangletools]

✅ Auto-approved drewstone PR — 68ad00eb

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-08T00:04:16Z}