test(wave-4): sybil-flood cohort #1 — coordinated-broadcast (W4.1, #196)

[Dewinator]

Summary

Third row of the W4.1 adversarial corpus, stacked on #198 (plagiarism row).

A canonical M=10 §10.4 cross-peer concentration cohort: ten signature-valid envelopes from ten distinct sock-puppet fixture-key node IDs, near-identical embeddings (cosine > 0.95), identical content + evidence_root, signed within a 5-minute window. Each envelope individually passes validateWireRecord — §10.4 is post-admission, REM-cycle, not admission-time.

The test asserts scoreFinding clamps local_weight strictly below MIN_LOCAL_WEIGHT_FOR_BROADCAST (0.3) at M=10 cohort size for both:

• Saturation: over_concentration=1.0 → new_weight=0
• Boundary: over_concentration=0.8 (the §10.4 trigger threshold) → new_weight=0.2 ≤ 0.3

Asserting at M=10 (vs plagiarism's M=3) proves §10.4 has no M-dependent escape hatch — the firebreak is purely a function of over_concentration, not of cohort size.

§10.2 quarantine arm — flagged, not silently weakened

The spec lists this row as "§10.2 quarantine + §10.4". §10.2 quarantines an origin only after N=5 consecutive rejections under §5 rules 1-20 (or single-strike on rules 5/16/19/20). A signature-valid well-formed sybil envelope passes those rules, so admission-time §10.2 does not fire on the cohort.

The §10.2 dimension of sybil-flood is therefore operator-driven multi-node escalation (the operator quarantines the M flagged origins after a §10.4 finding) and is deferred to W4.2 (HTTP corpus runner) and W4.3 (multi-node campaign report) per the issue #196 deferral rule. The deferral is captured in:

• metadata.owns_mechanism = "§10.2 quarantine + §10.4 diversity filter" so the spec linkage stays greppable
• metadata.comment — explicit text identifying the §10.2 arm as W4.2/W4.3
• The metadata-invariant test which asserts both spec sections appear in owns_mechanism

This satisfies the issue #196 acceptance criterion: "If a §10 mechanism cannot be made to fire from a unit test (because it requires real federation or a multi-node cohort), that fixture is deferred to W4.2 — flag it with a comment, do not weaken the assertion."

Stacking

• Base: agent/wave-4-plagiarism-fixture (PR test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198) → which is stacked on test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197
• Once test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198 lands, GitHub will auto-rebase this PR onto main

Files

• cohort-keys-sybil-flood.json — 10× Ed25519 fixture-only keys (write-once, same discipline as fixture-key.json and cohort-keys-plagiarism.json)
• sybil-flood/coordinated-broadcast-cohort.json — the M=10 cohort fixture
• anti-echo-sybil-flood.test.ts — 10 assertions across metadata, distinct-origin invariant, Pillar-6 signatures, wire-validator admission, §10.4 trigger conditions (cosine + signed_at window + shared evidence_root), and the scoreFinding decision layer

Test plan

• npx tsx --test src/__tests__/anti-echo-sybil-flood.test.ts — 10/10 green
• npm test — 963 pass / 0 fail (10 new sybil-flood tests + 953 prior)
• npm run build clean
• node scripts/build-anti-echo-fixtures.mjs is idempotent (re-run produces byte-identical fixtures, no working-tree drift)

Decomposition status

This closes row 3 of issue #196's 8-row table. 5 categories remain (echo-chamber, slow-poisoning, truth-by-repetition, polarity-inversion-pair, chain-rewrite). Each will land as a separate one-PR-per-row.

🤖 Generated with Claude Code