Epic: depth-axis parity for a multi-actor, shared-state simulation (generic capabilities)

[danielgwilson]

Summary

Epic / tracking issue for the depth-axis goal: make mimetic capable of replacing a bespoke multi-actor, shared-state product simulation with mimetic core + a thin, declared in-repo extension — with decision-equivalent evidence (docs/goals/proof-roadmap/goal.md). This is the hardest of the three depth targets because it stacks two capabilities mimetic does not yet have: a multi-persona swarm and a shared mutable service plane.

We have shipped the enabling layers (subject provider + subject.state seed/migrate, the scripted-browser actor, the terminal-product real-agent lane with command-scoped credentials + cost/no-spend ledger + the extension seam). We have not yet completed a single end-to-end bespoke-sim deletion proof. This epic tracks the remaining generic capabilities; the domain-specific adapter that consumes them is intentionally downstream / out of this repo (it is product-specific and stays in the adopter's own repo as the thin extension — that split is the extension-budget thesis).

Tracked capabilities (generic mimetic core)

In dependency order:

1. Multi-lane fan-out (layer 2) — Multi-lane fan-out for the computer-use lab (proof-roadmap layer 2) #163. The prerequisite: N differentiated persona lanes (per-lane persona/device/starting-surface), bounded concurrency, per-lane-worlds default. Was designed + judged previously; lands as its own goal-packeted phase.
2. Per-lane / per-actor starting surfaces — folded into Multi-lane fan-out for the computer-use lab (proof-roadmap layer 2) #163's roster (each lane enters at its own declared loopback route).
3. Multi-actor shared-world topology (layer 7) — Multi-actor shared-world (shared-service) topology (proof-roadmap layer 7) #164. The deep one: N actor roles against one provisioned mutable service plane; a declared override of per-lane-worlds whose open question is decision-grade attribution in a shared world (design-first).
4. Adapter scoring/feedback hook on the computer-use route — Adapter scoring/feedback hook on the computer-use (browser) route #165. Generalizes the terminal-lane extension seam so a domain rubric attaches without forking core.
5. First-class PII/PHI detector — Add a first-class PII/PHI detector to the redaction gate #108. Becomes load-bearing for any data-sensitive adopter; the deterministic shape tier first.
6. Live-rung prerequisites — Live terminal-product receipt: needs a dedicated test E2B key + a codex-bearing sandbox image (#154 SLICE 2 follow-up) #159. A dedicated, isolated test E2B key + a sandbox image with the agent runtime, so multi-lane/agent live receipts can be kept.

Recommended path (build the bridge, not a rip-out)

1. Prove the "delete-for-mimetic + thin adapter" thesis end-to-end on an easier single-actor target first, so the pattern is proven before betting it on a multi-actor shared-state target.
2. Build the generic capabilities above in order (fan-out → shared-world design+build → scorer hook → PII/PHI tier).
3. Keep any bespoke harness as the acceptance baseline until per-capability parity is demonstrated against kept mimetic run-bundle receipts.

Each capability ships as its own ratified goal packet + proof ladder (deterministic merge gate, one kept live rung). Do not rip out a working bespoke harness before parity.

Public-safety stance

Generic roadmap; no product names, domains, paths, or values. The domain adapter that consumes these capabilities is downstream and not tracked here.