chore(release): prepare agent-knowledge 6.1.11

[drewstone]

Updates the exact agent-eval dependency to 0.135.2 so memory promotion uses the corrected shared paired decision.

The memory improvement test data now contains varied positive gains instead of a constant sample, exercising statistical evidence without changing production policy.

Proof:

• 468 tests passed, 12 skipped
• 11 memory improvement tests passed
• lint, source and contract typechecks, build, and package verification passed
• official GEPA and SkillOpt integration checks passed
• production dependency audit found 0 known vulnerabilities

[tangletools]

✅ Auto-approved drewstone PR — 72b182ad

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 · 2026-07-29T06:24:18Z}

[tangletools]

🟢 Value Audit — sound

Verdict	sound
Concerns	0 (none)
Heuristic	0.0s
Duplication	0.0s
Interrogation	154.4s (2 bridge agents)
Total	154.4s

💰 Value — sound

Bumps agent-eval to pick up the corrected paired-promotion decision and diversifies the memory-improvement fixture so the significance bootstrap finally sees real variance instead of a degenerate constant — a coherent, minimal release-prep change.

• What it does: Two things. (1) Pins @tangle-network/agent-eval from 0.135.1 to 0.135.2 (package.json:78), which ships a corrected paired holdout decision used by decidePromotion in src/memory/improvement/promotion.ts:35. (2) Rewrites the final-sequence fixture in tests/memory/improvement.test.ts so the six final sequences carry varying same-agent probe counts ([0,1,2,3,1,2] at line 532) instead of each having a
• Goals it achieves: Make memory promotion consume the fixed paired-decision logic, and — critically — stop feeding that logic degenerate input. Before this PR every final cell was structurally identical (one delivery probe), so all six baseline cells scored exactly 0.25 and all six winner cells scored 1.0; the per-pair delta was a constant 0.75, giving heldoutSignificance's bootstrap resampler zero variance to work w
• Assessment: Good change on its merits. The dependency move is a patch bump following the repo's exact-pin convention (package.json:78, no caret), consistent with the CHANGELOG entry (CHANGELOG.md:3-8). The fixture refactor is clean: launchDateProbe is a pure extractor, sameAgentProbeCount is a backwards-compatible default-zero parameter, train/selection sequences are unaffected (they pass no third arg, line 4
• Better / existing approach: none — this is the right approach. I checked for an existing varied-probe fixture or a shared significance-variance helper to reuse (grepped tests/ and src/memory/experiment for probe/score/significance patterns); the improvement.test.ts fixture is the established place for this coverage and extending it is in-grain. A dedicated standalone significance-variance test would be scope creep for a rele
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 2
• Bridge warning: opencode/kimi-for-coding/k2p7: bridge stream ended without value-audit content

🎯 Usefulness — sound

Clean release-prep: bumps agent-eval to 0.135.2 so promotion flows through the corrected shared paired-decision, and replaces constant test data with varied positive gains that actually exercise the statistical gate.

• Integration: Fully reachable. The dep fix lands in heldoutSignificance (imported at src/memory/improvement/promotion.ts:2, invoked at :35 and :49), which drives both the promote/hold verdict and per-dimension regression checks. runAgentMemoryImprovement is a public export (src/memory/improvement → src/memory/index.ts:6 → the package's ./memory subpath), documented in AGENTS.md as one of the four canonica
• Fit with existing patterns: Matches the established release cadence exactly (git log shows chore(release) 6.1.6 → 6.1.11 as a repeating pattern). The test-data change aligns with how the codebase actually scores: promotion.ts:22-23 computes per-pair means and promotion.ts:35 runs a bootstrap significance test, so the prior constant-gain fixture (all 6 final sequences structurally identical → trivially 'significant') was a we
• Real-world viability: Holds up. The varied probe counts create non-uniform per-cell scores, so the significance path is now tested against realistic variance rather than a degenerate constant. The dependency bump is the standard mechanism for propagating a shared-primitive fix (the paired-decision lives in agent-eval and is consumed here unchanged); promotion.ts's own scoring is deterministic mean/lift math. No concurr
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 1

No concerns — sound change, no better or existing approach found. ✅

---

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260729T062804Z}