refactor: make experiment telemetry swarm-native

[christopherjnelson]

PR 3 of the zero-swarm migration. Follows #66.

Makes the export/archive telemetry pipeline swarm-native and removes the last legacy compatibility shims.

Export schema v12

schemaVersion is now z.literal(12); versions 9, 10 and 11 are dropped. Those versions existed to describe exports written before independent attempt accounting and, earlier, tick summaries — shapes the exporter can no longer produce.

• Attempt accounting becomes an unconditional invariant instead of a schema-v11-only rule. The paired conditionals ("schema-v11 exports require independent attempt accounting" / "Legacy exports cannot claim schema-v11 attempt accounting") collapse into one.
• ExperimentSource makes swarmTicks, providerAttempts, attemptRetention and attemptAccounting required; the version-conditional spreads and their ?? [] / ?. guards are gone.
• The archive importer accepts v12 only and rejects anything else outright. New test covers that rejection.

Last compatibility shim removed

archivedAppliedScenarioSchema no longer preprocesses cognitionMode/decisionContractVersion into historicalCognitionMode/historicalDecisionContractVersion. That removes the final legacy-multi-agent literal — grep -rn "legacy-multi-agent" apps packages now returns zero hits.

The schema keeps its name rather than collapsing into appliedScenarioSchema: it still tolerates archived scenarios written under relaxed pre-v2 execution-limit and objective defaults, which is unrelated to cognition architecture. A parallel manifest-level shim was removed alongside it.

Documentation

ADR 0033 records the retirement. It is deliberately explicit that the argument is structural — one planning call per tick versus one per agent per tick, which follows from the call pattern regardless of model or price — and states plainly that no run in this repository has compared legacy against zero-swarm cognition on real providers. The offline fixture was scripted with latency zero by construction and omitted cost; the live comparison is a Jev ablation that deliberately excluded legacy mode. The ADR claims no evidence it does not have.

Eleven ADRs specific to the retired architecture are marked Superseded by ADR 0033 in their own existing status style. Where supersession is partial (0009 model catalog, 0028 reflex seam, 0030 World Lab) the note says precisely which part is superseded and which still applies. Nothing is deleted or rewritten.

ZERO_SWARM_COMPARISON.md is marked a historical migration experiment with its findings, tables and numbers left intact. LIVE_SWARM_COMPARISON.md and EXPERIMENT_ARCHIVE.md get factual corrections only.

Comparison tooling

Verified, no changes needed — both harnesses, their CLIs, help text and tests already compare only Zero+Jev against Zero+deterministic workers. PR 1 had retired the legacy variant.

Validation

pnpm validate → pass. pnpm test:e2e → 2/2 pass. 221 tests across 26 files.

Known follow-ups

• EXPERIMENT_ARCHIVE.md is staler than its schema-version claims — its command list and storage-schema prose still describe tables dropped in PR 1/2. Flagged in-file; belongs to PR 5's documentation rewrite.
• world-lab.tsx still carries const swarmMode = true with dead !swarmMode branches — PR 4.

🤖 Generated with Claude Code