B4: harness-reframe prototype (experiment iteration → score + deterministic driver)

[aarontrowbridge]

Part of #109. Prototype for review — not a full close of the issue.

⚠️ This is the deepest / most-architectural slice of the depth-1 redesign and is deliberately shipped as a PROTOTYPE for review, not a merge-now change. It proves the shape of dissolving the LLM orchestrator for one experiment iteration; the full develop-mode DAG walk and retiring orchestrator.md are the B8 epic (#115), explicitly out of scope here.

What this proves

One experiment iteration runs end-to-end driven by a score + deterministic TS driver, not an LLM orchestrator:

select target (from ITERATION.toml, data)      ← no LLM
      │
dispatch ONE experimenter leaf (flat/depth-1)   ← the ONLY model seam; it authors the Julia script
      │
derive SolveSpec from the score + launch via    ← amico-run --spec runs the launch gate
the amico-run CLI (the §7.3 "harness calls         and, for tier-2/3, the independent re-rollout
the CLI directly" spine)
      │
record outcome + gate promotion on `agree`      ← bookkeeping, in code

The trust-tier, source exemplar, env, and promote decision are all data or code — the only judgment call in the loop is the single flat leaf that authors the script. That is the orchestrator dissolving, for one iteration.

Stands on #107 (G-1)

Per #107's recorded ruling: score-first for flow + a thin TS driver only for what the stage model can't express. The experiment-iteration loop is exactly that exception, so the flow lives as data (an ITERATION.toml harness score) and the loop lives as a thin driver in the amico-run (CLI) lineage. Blocked-by #107.

Acceptance criteria (#109)

• one experiment iteration runs end-to-end driven by score + driver, not an LLM orchestrator
• experimenter dispatched flat (depth-1); re-rollout gate runs incl. tier-2 (composed); promotion gated on agree
• no control-flow logic left in an agent prompt for this path (it is runExperimentIteration, TS)

What changed

• packages/amico-run/src/harness/ — iteration_score.ts (parse/validate the data-defined flow) + experiment_iteration.ts (the deterministic loop; the single model seam dispatchExperimenter is pluggable — production wires it to opencode run --agent experimenter).
• Re-rollout gate extended to tier-2 (composed) per spec §4.3, not just tier-3/free — isVerifiedTier() in verify.ts is the single source of truth; cli.ts gates on it.
• packages/extension/scores/experiment-iteration/ITERATION.toml — the example harness score (data). It is not a SCORE.md, so the interview repertoire loader ignores it by design.
• Runnable demo + tests with deterministic fakes (no LLM, no Julia).
• s31 grep-guard now recurses src/ so it also covers the new harness subdir (tool layer stays CLI/spawn: no fetch/MCP/HTTP).

What a reviewer runs

pnpm install
pnpm --filter @amicode/schema build
pnpm --filter @amicode/amico-run build
pnpm --filter @amicode/amico-run test          # 115 pass (was 106); the B4 acceptance is experiment_iteration.test.ts
pnpm --filter @amicode/amico-run demo:harness   # prints each deterministic step + the outcome, ends "PASS … No LLM in the control-flow loop."

Verified locally: amico-run typecheck clean, 115/115 tests pass, demo prints PASS. Extension scores/lint tests unaffected (the new dir is ignored by the interview loader).

Design choices a human should confirm

1. Driver home = amico-run (CLI lineage), not the VS Code extension. Rationale: spec §7.3 makes the deterministic harness "call the CLI directly — no LLM," and the driver shells amico-run --spec so the launch gate + re-rollout are reused, not re-implemented. The example score still lives under extension/scores/. If reviewers want the driver in the extension proper, that's a move, not a rewrite.
2. The SolveSpec is derived from the score, not authored by the leaf. The leaf authors only the Julia script; tier/source/env come from the data. This maximizes the "no LLM in control flow" property. Confirm that's the intended division of labor for the experimenter contract.
3. Promotion policy for non-verified tiers. promote_on = "agree" gates verified tiers on the re-rollout; a vetted (tier-1) target has no re-rollout, so the driver promotes on completed (template-trusted). Confirm that's the desired tier-1 behavior.
4. isVerifiedTier = {free, composed}. This is the §4.3 tier-2 extension (G-7 leaned default-on). Confirm default-on for composed vs per-score opt-in.

Notes

• Fork CI may be RED — pre-existing cross-repo token blocker (vendored-binary fetch), unrelated to this change.
• Do not merge; this is up for review on top of a working B1 and for Jack's input, per the B0 (decision, Jack): ratify depth-1 redesign gates G-1/G-8/G-9 #107 plan.

🤖 Generated with Claude Code