feat(wave-4): W4.1 — fill the anti-echo corpus (8 fixtures + node:test harness)

[Dewinator]

Scaffold landed in commit f52a914:

• mcp-server/src/__tests__/fixtures/anti-echo/README.md
• mcp-server/src/__tests__/fixtures/anti-echo/corpus-types.ts

Per docs/wave-4-anti-echo.md §"Sub-task decomposition", W4.1 is the first piece of Wave 4 and is explicitly agent-eligible. This issue tracks the work after the scaffold: fill the corpus and add the harness, one category at a time so each PR stays small and reviewable.

Decomposition (one PR per row, in any order)

Each PR adds:

1. A <category>/ subdirectory with its first JSON fixture (or cohort).
2. A node --test test file (anti-echo-<category>.test.ts) under mcp-server/src/__tests__/ that loads the fixture via AntiEchoCorpusFixture/AntiEchoCohortFixture and asserts the §10 mechanism rejects/quarantines/demotes per metadata.expected_outcome.
3. If the PR is the first to need a signing key, also lands fixture-key.json (Ed25519, fixture-only, deterministic seed). The README already reserves the file slot.

#	Category	§10 mechanism it owns	Single envelope or cohort?
1	forgery	§3.7 PoK + §5 rule 1	single
2	plagiarism	§10.4 diversity filter	cohort (N≥3, same key OR different keys depending on §10.4 sub-rule the test exercises)
3	sybil-flood	§10.2 quarantine + §10.4	cohort (M≥10)
4	echo-chamber	§10.4	cohort (≥80% of synthetic peer set)
5	slow-poisoning	§10.5 REM self-audit	single (paired with a high-confidence local experience the test seeds first)
6	truth-by-repetition	§10.4 + §10.6 firebreak	cohort (N keys, asserts no Tier-A promotion)
7	polarity-inversion-pair	§10.3 contradicts-trigger	cohort (size 2)
8	chain-rewrite	§5 rule 19 (single-strike)	single

Acceptance criteria

• All 8 categories have at least one fixture and one test case.
• Every assertion checks both the rejection (admission decision) and the side effect (trust-edge delta, nodes.quarantined_until, memory_relations.contradicts edge, REM falsification reason) — per wave-4-anti-echo.md §"Corpus categories".
• Tests live alongside the existing lesson-admission.test.ts, rem-audit.test.ts, swarm-admit.test.ts etc. and pass with npm test.
• No fixture changes the production code path. 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 (HTTP corpus runner, separate issue) — flag it with a comment, do not weaken the assertion.

Out of scope for this issue (covered later)

• W4.2: scripts/run-anti-echo-campaign.mjs HTTP runner against a self-loop federation listener.
• W4.3: multi-node campaign + report generator → docs/constitution-defense-report.md. Operator-driven, blocked on Wave 2.

Related

• Anchor doc: docs/wave-4-anti-echo.md
• Wave overview: docs/waves.md
• Spec: docs/SWARM_SPEC.md §10
• Scaffold commit: f52a914

[Dewinator]

W4.1 progress checkpoint — paused at 2/8 to avoid queue saturation

Progress against the decomposition table:

• ✅ Row 1 — forgery/evidence-root-mismatch → PR test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197 (CI green, MERGEABLE)
• ✅ Row 2 — plagiarism/identical-broadcast-cohort → PR test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198 (CI green, MERGEABLE, stacked on test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197)

Remaining: rows 3–8 (sybil-flood, echo-chamber, slow-poisoning, truth-by-repetition, polarity-inversion-pair, chain-rewrite).

Why pause now

Queue is currently 11 PRs deep — 9 native-app-track PRs (#185, #187–194) + #197/#198. All green, all MERGEABLE. Two soul-traits point the same way:

1. "Wenn die PR-Queue voll und auf Reed wartet, dann Merge-Order analysieren — nicht neue PRs erstellen" (evidence 21).
2. Form-saturation rule — after ~5 ticks of the same shape (here: docs-on-main + fixture-row PRs), break form before the next.

Shipping rows 3–8 in sequence right now would push the queue to 17 deep without changing what blocks Reed.

Orthogonality assurance (the pause is queue-hygiene, not a merge risk)

• test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197 and test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198 only touch new files under mcp-server/src/__tests__/fixtures/anti-echo/, mcp-server/scripts/build-anti-echo-fixtures.mjs, mcp-server/src/__tests__/anti-echo-*.test.ts, and docs/wave-4-anti-echo.md. Zero file overlap with any of the 9 native-app-track PRs.
• Within the pair: test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198 stacks on test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197 (shared scaffold). Merge order: test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197 → test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198, or test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198 alone via GitHub's auto-rebase.
• Lockfile risk lives only inside the native-app stack (feat(native): PGlite adapter foundation — issue #184 part 1 #185 vs feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1 #187 both touch mcp-server/package.json); Wave-4 fixtures touch no package files.

Resumption trigger

Pick this issue back up for rows 3–8 once any of the 9 native-app-track PRs lands. Each remaining row is one focused PR matching the established pattern: fixtures/anti-echo/<category>/*.json + anti-echo-<category>.test.ts. Estimated effort per row: 1 tick.

🤖 Posted by autonomous tick (mycelium 127th)

[Dewinator]

Per-row triage shipped — all 6 remaining rows are unit-testable on main

Closed the implicit "is this row deferrable to W4.2?" question per row before the next agent picks W4.1 back up. New doc on main: docs/wave-4-w4.1-row-triage.md (commit cbf85b4).

Row	Category	§10 mechanism	Existing unit-test path on main	Verdict
3	sybil-flood	§10.2 + §10.4	swarm-admit.test.ts + rem-diversity.test.ts	unit-testable
4	echo-chamber	§10.4	rem-diversity.test.ts	unit-testable
5	slow-poisoning	§10.5	rem-audit.test.ts	unit-testable
6	truth-by-repetition	§10.4 + §10.6	rem-promotion.test.ts	unit-testable
7	polarity-inversion-pair	§10.3	lesson-contradiction-gate.test.ts	unit-testable
8	chain-rewrite	§5 rule 19	lesson-chain.test.ts	unit-testable

No row needs a // TODO: deferred to W4.2 comment. No row weakens its assertion. When the resumption trigger fires (any of the 9 native-app PRs lands), each row is one focused PR matching the established pattern from #197 / #198 — fixture JSON + anti-echo-<category>.test.ts against the existing dep-stub harness.

Pause at 2/8 still holds — this is queue-prep, not a row-3 PR.

🤖 Posted by autonomous tick (mycelium 132nd)

[Dewinator]

Row-triage doc fact-checked against main (this tick, 2026-05-03)

Spot-checked every claim in docs/wave-4-w4.1-row-triage.md against the actual code on main. All 6 hold:

Row	Claim	Verifier on main	Result
3	swarm-admit.test.ts mints peer keys via generateKeyPairSync and signs envelopes with signWithSelfKey	grep generateKeyPairSync|signWithSelfKey mcp-server/src/__tests__/swarm-admit.test.ts	✅ both present
4	rem-diversity.test.ts exercises scoreFinding directly	grep — 15 occurrences	✅
5	rem-audit.test.ts follows the "pure-helper + orchestrator with Deps stubs" pattern	header lines 7–32 explicitly document this; RemAuditDeps interface present at line 8	✅
6	rem-promotion.test.ts exercises scoreFinding against synthetic clusters	grep — 9 occurrences	✅
7	lesson-contradiction-gate.test.ts exercises cosineSimilarity + polarityInverted in isolation	header lines 7–24 document the "cosine > 0.85 AND polarity inverted" gate	✅
8	lesson-chain.test.ts exercises detectLessonChainConflict against prev_lesson_hash	grep — both symbols present, including the Postgres-23505 mapping tests at lines 183/194	✅

What this means: when the queue drains and the next agent picks row 3 (sybil-flood), the doc's "pattern: copy swarm-admit.test.ts shape, mint M ≥ 10 keys, assert quarantine" is grounded in real code on main, not aspirational. No mid-tick discovery that the claimed harness doesn't exist.

W4.1 stays paused at 2/8 per the prior commitment — this is verification, not progress.

[Dewinator]

Local merge-validation snapshot (149th tick, 2026-05-03): stacked PRs #197 → #198 build clean and pass empirically end-to-end.

Method: fetched both PRs as local branches, checked out PR #198 (which carries the full stack on top of main), rm -rf dist && npm run build clean, then ran node:test against the built artifacts.

Results on the #197+#198 stack:

• Full mcp-server suite: 953 / 953 pass, 0 fail, 0 skip (6.7 s).
• W4.1-specific (anti-echo-forgery.test.js + anti-echo-plagiarism.test.js): 14 / 14 pass (5 forgery, 9 plagiarism). The §10.4 boundary case (scoreFinding at over_concentration=0.8 still clamps below broadcast threshold) and the saturation case (= 1.0 silences) both fire as the issue's acceptance criteria require — both rejection AND side-effect (broadcast-suppression score) are asserted.
• Build script idempotency: node scripts/build-anti-echo-fixtures.mjs re-run on top of an already-populated tree → exits 0, git status empty. The PR test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197 body's fixture-key.json "refused-overwrite on every subsequent run" claim holds in practice.

Implication for merge order: these two are independent of the wave-1 9-PR native-app stack and the wave-2 fail-soft PR #199. They can land in any of {#197 → #198}, {wave-1-stack}, {#199, #200} order without affecting each other — the 13-PR queue stays merge-independent (consistent with the 148th-tick proof).

Remaining W4.1 rows (unit-testable, per commit cbf85b4): sybil-flood, echo-chamber, slow-poisoning, truth-by-repetition, polarity-inversion-pair, chain-rewrite. Following PRs in the chain stack on #198 the same way #198 stacks on #197 (shared fixture-key.json + build-anti-echo-fixtures.mjs).

[Dewinator]

W4.1 progress checkpoint — 3/8 (sybil-flood shipped, paused at 3/8)

Updates the prior "paused at 2/8" comment (127th-tick) with row-3 progress shipped in the 151st tick.

Decomposition table — current state

• ✅ Row 1 — forgery/evidence-root-mismatch → PR test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197
• ✅ Row 2 — plagiarism/identical-broadcast-cohort → PR test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198 (stacked on test(wave-4): forgery fixture #1 — evidence-root-mismatch (W4.1, #196) #197)
• ✅ Row 3 — sybil-flood/coordinated-broadcast-cohort → PR test(wave-4): sybil-flood cohort #1 — coordinated-broadcast (W4.1, #196) #201 (stacked on test(wave-4): plagiarism cohort #1 — identical-broadcast (W4.1, #196) #198) — M=10 cohort, 10 distinct fixture-key Ed25519 IDs, asserts §10.4 over_concentration clamps local_weight strictly below MIN_LOCAL_WEIGHT_FOR_BROADCAST at both saturation (1.0) and §10.4-trigger boundary (0.8)
• ⏳ Rows 4–8 — echo-chamber, slow-poisoning, truth-by-repetition, polarity-inversion-pair, chain-rewrite (all unit-testable per docs/wave-4-w4.1-row-triage.md)

Stack health (this tick, 2026-05-03)

PR	mergeable	mergeStateStatus	CI	Stack base
#197	MERGEABLE	CLEAN	SUCCESS	f52a914 (main scaffold)
#198	MERGEABLE	CLEAN	SUCCESS	5aafa86 (= #197 head)
#201	MERGEABLE	CLEAN	SUCCESS	4b4819b (= #198 head)

Linear single-commit stack: f52a914 ← 5aafa86 ← 4b4819b ← 82d8b79. Each PR has exactly 1 commit. Reed can gh pr merge in stack order (#197 → #198 → #201) without manual rebase.

Main-has-moved orthogonality

Main has moved 5 commits ahead of #197's base (f52a914 → d7a0710) since the W4.1 stack opened:

• cbf85b4 docs(wave-4): W4.1 row triage
• f9444e3 docs(wave-4): correct W4.1 test runner
• 6706161 docs(wave-2): flag mTLS-listener build-deferred drift
• 65cd5ed docs(wave-2): cross-link PR fix(federation): null-guard mTLS listener when deferred build missing — W2 follow-up to #6706161 #199
• d7a0710 docs(wave-3): anchor doc for tracker-free discovery

All 5 are docs/-only. Diff between f52a914 and d7a0710 shows zero overlap with the W4.1 fixture/test paths under mcp-server/src/__tests__/anti-echo-*.test.ts, mcp-server/src/__tests__/fixtures/anti-echo/**, or mcp-server/scripts/build-anti-echo-fixtures.mjs. GitHub's CLEAN mergeStateStatus on all 3 PRs corroborates: the rebase delta is doc-only.

Pause discipline still holds

Queue is now 14 PRs deep (9 native-app-track + #197/#198/#199/#200/#201). Same trait-1 reasoning as the 127th-tick pause comment applies: shipping rows 4–8 in sequence right now would extend the queue to 19 without changing what blocks Reed. Each remaining row stays one focused PR matching the established #197/#198/#201 pattern (one new test file + one fixture JSON + a build-script edit).

Resumption trigger (unchanged)

Pick row 4 (echo-chamber) up once any of the queue's 14 PRs merges. The pattern: stack onto the latest of #201 or post-#201 W4.1 PR, mint a fresh cohort-keys-echo-chamber.json (Ed25519 sock-puppets), exercise rem-diversity.test.ts's scoreFinding against an ≥80% near-identical synthetic peer set.

🤖 Posted by autonomous tick (mycelium 152nd)