feat(runtime): MAN graph rewrite training harness

docs/07-mvp/morphogenetic-paper-validation.md

@@ -0,0 +1,196 @@
+# Morphogenetic paper validation (§7–§8)
+
+Two layers (paper §8):
+
+| Layer | Hypothesis | Status | Entry |
+| --- | --- | --- | --- |
+| **Phase I** | H1–H5 (+ F1–F3 foundations) | **Done** on fixtures | `bash scripts/run-man-paper-experiments.sh` |
+| **Phase II** | **H0** self-evolution capability | **Runnable** (coding + OpenClaw stub) | `bash scripts/run-man-paper-phase-ii.sh` |
+
+Phase I proves mechanisms per spec; passing Phase I is **necessary but not sufficient** for H0.
+Phase II will show learning curves (no mid-run dev edits), A→B/cross-domain transfer, and
+MAN vs static vs developer-effort-matched hand-iterated baselines on coding/OpenClaw scenarios.
+
+---
+
+## Phase I authenticity (what is “real” vs proxy)
+
+| ID | Data source | Real runtime path? | Known proxy / caveat |
+| --- | --- | --- | --- |
+| **H1** | `h1_longitudinal.csv` | **Yes** — `EffectorKernel.run_turn`, `lifecycle.reinforce` on success, `guard_score` = `concern.score`; warm `step_delta≈0`, no cold in H1 epoch | CPS = planner + fractional verifier, **not** billed LLM tokens |
+| **H2** | `r_t_bimodal` + `r_t_bandit` JSONL | **Yes** — `CreditField`, `evaluate_split_guards`, variance on buffer | No live agent loop; fixture replay only |
+| **H3 ρ** | single activation vector | **Yes** — `tier1_responsibility` math | Not a multi-episode run |
+| **H3 plasticity** | bandit JSONL | **Yes** — `split_with_spec_or_keywords` | Parent **primed** with `reinforce(δ=0.15)` so cold split is eligible (fixture scaffolding) |
+| **H4 ρ** | synthetic hard/soft pair | **Yes** — tier-1 ρ | Not stochastic LLM |
+| **H4 noise** | simulated Bernoulli outcomes | **No** — Monte Carlo outcome noise | ρ gap real; outcome variance **simulated** |
+| **H5 soak** | `r_t_soak_long.jsonl` | **Yes** — append RT + `warm`/`cold` plasticity | Was wrongly on 32-row bimodal only — **fixed** to soak file |
+| **F1–F3** | bimodal JSONL | **Yes** — replay / λ / β on buffer | — |
+| **Main table** | demo-tool 40 turns | Partial | Illustrative baselines only |
+
+**Not Phase I:** application Self-Evolving (H0), real token curves, hand-iterated baseline.
+
+---
+
+## Phase II (H0)
+
+```bash
+bash scripts/run-man-paper-phase-ii.sh
+```
+
+Runs Phase I first, then **H0 genesis** only (`experiments/man_paper/phase_ii_seed.py` —
+see [`docs/design/h0-genesis.md`](../design/h0-genesis.md)): startup prompt → one
+`intent_alignment` **cortex** concern + optional conserved reflex `h0.conserved.fail-closed`.
+**No** plugin `seed_stores()`, **no** `SKILL.md` concern init, **no** demo coding/OpenClaw
+presets. Cross-domain scenarios use the same H0 graph on `before_response` (tool-style cases
+judged from LLM refusal text when effector path is not exercised).
+**LLM:** auto-detect from env (**B.AI first**, same order as daemon `provider: auto`):
+
+```bash
+# B.AI (recommended on this machine)
+export BAI_API_KEY='sk-...'
+export BAI_MODEL='gpt-5.2'   # optional; default gpt-5.2
+uv run python experiments/man_paper/phase_ii_run.py --epochs 20
+
+# Or explicit
+uv run python experiments/man_paper/phase_ii_run.py --provider bai --epochs 20
+
+# Other providers
+export OPENAI_API_KEY=...
+uv run python experiments/man_paper/phase_ii_run.py --provider openai --epochs 20
+
+# Load from ~/.opencoat/opencoat.env (if BAI_API_KEY is there)
+set -a && source ~/.opencoat/opencoat.env && set +a
+uv run python experiments/man_paper/phase_ii_run.py --epochs 20
+
+# CI / offline: prompt-aware stub
+OPENCOAT_PHASE_II_FORCE_STUB=1 uv run python experiments/man_paper/phase_ii_run.py --provider stub
+```
+
+See [`docs/config/bai-llm.md`](../config/bai-llm.md).
+
+With a **real** LLM, gates use the `real_llm_advisory` profile (exit 0 unless `--strict-gates`).
+Stub mode keeps strict gates for CI.
+
+| Output | Content |
+| --- | --- |
+| `results/PHASE_II_RESULTS.md` | Learning curves + transfer + H0 gates |
+| `results/phase_ii_learning_curves.csv` | `mode, epoch, success_rate, …` |
+| `results/phase_ii_report.json` | Full report |
+
+**Clean H0:** no split-gate priming — score/activations/reinforced state come only from
+`r − baseline` via `turn_complete` warm reweight + credit buffer; cold split enabled; see
+[`h0-genesis.md`](../design/h0-genesis.md). Phase I H3 still uses primed fixtures for
+*mechanism* validity only.
+
+**Before B.AI (budget):**
+
+1. Stub precheck (free): `OPENCOAT_PHASE_II_FORCE_STUB=1 uv run python experiments/man_paper/phase_ii_run.py --epochs 8` — confirm `first_split_epoch` and `split_guard_reason` evolve (expect split by ~epoch 4 with `split_n_min=24`).
+2. B.AI pilot: `--epochs 3` — watch `split_guard_reason` (plumbing vs signal vs denoise).
+3. Full run: `--epochs 20`. Contrast arm only: `--feature-mode text` (format drift).
+
+Baselines: **MAN** (single extracted concern + RT + lifecycle; structure grows via plasticity only),
+**static** (same bootstrap concern, frozen score, no plasticity), **hand_iterated** (bootstrap +
+hand patches on failure, dev budget 3). Transfer: `coding_heldout`, `openclaw_cross`.
+
+Bootstrap: [`docs/design/h0-genesis.md`](../design/h0-genesis.md) — *Start up. You are a
+Self-Evolving Agent…* → `extract_for_intent_alignment` (cortex) + `h0.conserved.fail-closed`
+(reflex core, `reflex: true`).
+
+Live OpenClaw gateway (product path, not H0 harness): daemon + bridge per
+`integrations/openclaw-opencoat-bridge/README.md` uses **plugin** genesis, not `seed_h0_graph`.
+
+---
+
+## Phase I: how to read the data
+
+Phase I is **not** a large benchmark. It is a **mechanism harness** with boolean gates on small,
+deterministic fixtures. Start here:
+
+| Read first | Why |
+| --- | --- |
+| `experiments/man_paper/results/INTERNAL_VALIDITY.md` | Scale table, pitfalls, gates, H1 summary, per-H metrics |
+| `experiments/man_paper/results/h1_longitudinal.csv` | Epoch × mode CPS/success (plot-friendly) |
+| `experiments/man_paper/results/internal_validity.json` | Machine-readable gates + `scale` block |
+| `RESULTS.md` | **Auxiliary only** — tiny demo-tool loop, easy to misread |
+
+### Fixture profiles (`fixtures/morphogenetic/scale.json`)
+
+| Profile | Bandit rows | Soak rows | H1 default |
+| --- | --- | --- | --- |
+| `standard` | 96 | 256 (8×32) | 20 epochs × 60 trials |
+| **`stress`** (default) | 384 | 1024 (32×32) | 20 epochs × 60 trials |
+
+Regenerate stress fixtures:
+
+```bash
+uv run python scripts/generate_morphogenetic_validation_data.py --scale stress
+```
+
+Legacy smaller set: `--scale standard`.
+
+### Four pitfalls (do not misread)
+
+1. **H1 `llm_calls_per_success`** — counts symbolic planner/verifier steps, not OpenAI tokens.
+2. **Main table in `RESULTS.md`** — ~40-turn demo; not the H1 learning curve.
+3. **H3** — often `tier1_splits=1` vs `uniform_splits=0`; not a large spurious-split study.
+4. **H5 soak** — bounded span on repeated bimodal replay; not Phase II competence growth.
+
+---
+
+## Generate fixtures
+
+```bash
+uv run python scripts/generate_morphogenetic_validation_data.py --scale stress
+```
+
+Writes JSONL under `packages/opencoat-runtime/tests/fixtures/morphogenetic/` plus `scale.json`.
+
+## Run validation
+
+```bash
+uv run pytest packages/opencoat-runtime/tests/integration/test_morphogenetic_paper_validation.py -v
+```
+
+## Paper §8 Phase I harness
+
+```bash
+bash scripts/run-man-paper-experiments.sh
+```
+
+Fixtures (stress) + mechanism pytest + `experiments/man_paper/run.py`. **Exit 1** if any
+Phase I gate (H1–H5, F1–F3) fails.
+
+Outputs:
+
+| Artifact | Content |
+| --- | --- |
+| `results/INTERNAL_VALIDITY.md` | **Primary** — scale, pitfalls, gates, metrics |
+| `results/h1_longitudinal.csv` | H1 curve data |
+| `results/internal_validity.json` | Gates + `scale` |
+| `results/PHASE_II_PROTOCOL.md` | H0 preregistration |
+| `results/report.json` | Full dump + auxiliary tables |
+| `results/RESULTS.md` | Demo baselines, sweeps (secondary) |
+
+Tests: `test_morphogenetic_internal_validity.py`, `test_man_paper_full_empirical.py`.
+
+## What is verified (mechanism unit tests)
+
+| Paper claim | Test |
+| --- | --- |
+| Credit conservation `Σκ_a ≈ r−b` | `test_credit_conservation_on_fixture` |
+| JSONL replay deterministic | `test_replay_deterministic_scores_and_edges` |
+| Split lowers child variance | `test_split_reduces_reward_variance` |
+| Eligibility trace `e ← λe + α·part` | `test_eligibility_trace_accumulates_and_decays` |
+| Tier-1 ρ weights hard > soft | `test_tier1_vs_uniform_responsibility_spread` |
+| Graph evolution from session | `test_rt_service_session_grows_connectome` |
+
+## Implementation map
+
+- `credit/eligibility.py` — `e_a`, `e_s`
+- `credit/baseline.py` — context bucket `b`
+- `credit/attribution.py` — tier-1 `ρ`
+- `credit/credit_field.py` — conserved κ + synapse ledger
+- `credit/synapse_ledger.py` — κ_s → edge LTP
+- `credit/tier2_calibration.py` — deterministic LOO on buffer
+- `credit/rt_replay.py` — full replay harness
+- `experiments/man_paper/` — Phase I harness (`internal_validity.py`, `phase_i_readme.py`)

docs/07-mvp/v0.3-delivery-status.md

@@ -0,0 +1,52 @@
+# v0.3 delivery status
+
+Tracks **M6 + v0.3 phase-2 full delivery** on `main` after the mega PR
+([#83](https://github.com/HyperdustLabs/OpenCOAT/pull/83)) and the follow-up
+**full TCB** PR (memory sync guards + message verify→repair).
+
+Re-run: `./scripts/verify-m6-prerequisites.sh` and bridge `npm test`.
+
+## Product scope (shippable)
+
+| Area | Status | Evidence |
+| --- | --- | --- |
+| M6 lifecycle workers | ✅ | `examples/07_meta_governance_soak`, pytest soak |
+| In-proc ReflexMonitor (6 guards) | ✅ | `tool` / `spawn` / `message_out` / `queue` / `memory_write` / `tool_result_persist` |
+| Message verify→repair | ✅ | `reflexMessageGuardDecision` + REWRITE export |
+| `r_t` JSONL + PlasticityEngine | ✅ | bridge `emitRtJsonl`, `credit.r_t.append`, warm reweight + cold lift/archive |
+| JSONL replay determinism | ✅ | `rt_replay.py`, `test_r_t_replay.py` |
+| **Connectome split (cold)** | ✅ | `split_spec.py` (H/G/ΔF) + `connectome_plasticity.py` |
+| **CreditField κ** | ✅ | `credit_field.py` + `rt_buffer.py` |
+| **Plasticity 6 primitives** | ✅ | warm: reweight/connect/prune; cold: split/lift/merge/archive |
+| **Architecture (ii) routing** | ✅ | `connectome/router.py` — synapse-weighted MoE + lift hub expansion |
+| **Architecture (ii) synapse evolution** | ✅ | `synapse_evolution.py` — edge LTP/LTD + co-activation strengthen |
+| **Weave co-activation → graph** | ✅ | `JoinpointPipeline` recorder → `RtPlasticityService.consume` |
+| **Paper κ / eligibility / baseline** | ✅ | `credit_field.py` + `eligibility.py` + `baseline.py` — tier-1 conservation |
+| **Paper validation suite** | ✅ | `test_morphogenetic_paper_validation.py` + fixture generator |
+| **Tier-2 calibration** | ✅ (LOO deterministic) | `tier2_calibration.py` — buffer leave-one-out |
+| **EffectorKernel.run_turn** | ✅ | weave → mediate → verifier repair → `r_t` |
+| **JSON-RPC** | ✅ | `effector.run_turn`, `plasticity.cold_step`, `credit.connectome.stats`, `connectome.route` |
+| Queue dogfood | ✅ | `examples/09_queue_hook_dogfood` |
+
+## Verification checklist
+
+```bash
+./scripts/verify-v0.3-full-spec.sh
+```
+
+## Explicitly deferred (ops / research)
+
+- **Tier-2** counterfactual ΔF calibration (scaffold only)
+- **Fork effector** multi-`propose` in one OpenClaw turn (daemon routing + graph evolution landed)
+- 24h live daemon soak (ops, not blocking merge)
+- PyPI release publish workflow — version bumped to 0.2.0; track packaging PR when publishing
+
+## Config reference
+
+OpenClaw plugin `@hyperdustlabs/opencoat-bridge`:
+
+- `inProcReflexToolGuard: true` — enables all in-proc guards including sync memory hooks
+- `reflexSyncFromDaemon: true` — pull `reflex.policies.export` on load
+- `emitRtJsonl: true` — append structured `r_t` to daemon
+
+See [bridge README](../../integrations/openclaw-opencoat-bridge/README.md).

docs/design/h0-genesis.md

@@ -0,0 +1,92 @@
+# H0 genesis (合子) — single contract
+
+Phase II / paper **H0** uses **one** genesis path. Product install and SkillLearnBench
+have **different** seeds; do not mix them into H0 attribution.
+
+## Three genesis paths in the repo
+
+| Path | Where | What gets seeded | H0? |
+| --- | --- | --- | --- |
+| **H0 experimental** | `experiments/man_paper/phase_ii_seed.py` | Startup prompt → `extract_for_intent_alignment` (cortex) + optional **conserved reflex** (brainstem) | **Yes — canonical** |
+| **Product / plugin** | `opencoat_runtime_cli/plugin_templates/*/bootstrap_opencoat.py` | `seed_concerns()` / `seed_stores()` — hand-authored plugin scaffold | **No** — install/dogfood only |
+| **Skill file injection** | SkillLearnBench `SKILL.md` → solving agent (or mistaken “Concern init”) | Pre-written skill text | **No for `man_full`** — use as **`skill_seed` baseline** only |
+
+## H0 graph at \(t=0\) (`seed_h0_graph`)
+
+After `seed_h0_graph()`:
+
+1. **Cortex (plastic)** — exactly **one** concern from:
+
+   > Start up. You are a Self-Evolving Agent via Reward-Modulated Structural Plasticity over an Externally Reasoning LLM.
+
+   - `source.origin = intent_alignment`
+   - `reflex: false` — subject to warm plasticity / lifecycle
+   - `pointcut`: `before_response`, `match: null` (joinpoint-only)
+
+2. **Conserved core (non-plastic)** — `h0.conserved.fail-closed`
+
+   - `reflex: true`, `neuron_type: inhibitory`
+   - Deterministic TOOL_GUARD on destructive shell patterns (`before_tool_call`)
+   - Excluded from `⇩_slow` structural rewrite (`PlasticityEngine` / connectome `is_conserved`)
+
+3. **Edges** — **0** at genesis (`dcn.add_node` only; co-activation / cold wiring comes later).
+
+This is **not** “pure single-cell with no brainstem”: it matches v0.3’s split between
+**A_reflex** (fail-closed invariant) and **excitatory cortex** (morphogenetic target).
+
+## What H0 explicitly does **not** use
+
+- `02_coding_agent_demo` / `04_openclaw_with_runtime` `seed_concerns()`
+- `demo-tool-block` or other Phase I fixtures
+- Plugin `bootstrap_opencoat.seed_stores()` presets
+- **SKILL.md → Concern upsert** as MAN’s initial graph (pollutes “what did morphogenesis grow?”)
+
+## SkillLearnBench alignment
+
+| Config | Genesis |
+| --- | --- |
+| **`man_full`** | H0 bare graph (`phase_ii_seed.seed_h0_graph`) inside runtime; skills are **exports** after learning (Mode A), not the seed |
+| **`static_aspect_graph`** | Same H0 seed, plasticity off |
+| **`hand_iterated`** | H0 seed + developer patches on failure (Phase II harness) |
+| **`skill_seed` / b1-style** | Pre-authored `SKILL.md` only — **separate baseline**, compare against `man_full` to quantify **seed vs grow** |
+
+See [`skilllearnbench-h0-integration-plan.md`](skilllearnbench-h0-integration-plan.md).
+
+## Clean H0 plasticity (no priming)
+
+Phase II **`man_full`** must **not** hand-prime split gates (contrast with Phase I H3 in
+`ablations.py`, which uses `reinforce(δ=0)×5` + `reinforce(δ=0.15)` so cold split is eligible on
+fixtures).
+
+| Gate | Organic source in H0 |
+| --- | --- |
+| `buffer ≥ 8` | `CreditField.attribute_turn` per woven turn (cortex in `active`) |
+| `state == reinforced` | First `lifecycle.reinforce` from warm `PlasticityEngine.reweight` on positive `r − b` |
+| `activations ≥ 3` | Three reinforce events (weaken does not count) |
+| `score ≥ 0.65` | Cumulative `min(|advantage|·step_delta, step_delta)` from successes — **bottleneck** |
+
+Harness: `experiments/man_paper/phase_ii_runner.py` — no bootstrap `reinforce(δ=0)`, no fixed
+`0.1/0.02` score bumps; `turn_complete` rows carry `active_aspects` + `reward` for cortex
+attribution; EMA baseline (`baseline_ema_alpha≈0.22`); **stable split axis**
+`payload["feature"] = scenario_id` (never LLM response text — see `credit_field._feature_from_record`);
+`split_spec` uses **categorical equality** or token axes (not `feat in long_string`);
+`split_n_min≥24`, **Welch** mean-gap test, **score EMA** for split eligibility;
+optional `rollout_k>1` averages `r` per context (pass^k denoise).
+Report **`first_split_epoch`**, **`last_split_guard_reason`**, and `h0_plasticity` in
+`phase_ii_report.json` (measured, not preset). Do **not** lower θ_sep / δ_min or prime scores to force split.
+
+**Lift vs split:** both use EMA score for gates; reflex lift runs only after `buffer ≥ n_min` and
+split guards decline (no `lift_score=0.96` shortcut). **β≈0.01** scales ΔF complexity to binary
+rewards (`G > 8β`); θ_sep and Welch carry the structural load.
+
+**Ablation (contrast arm only):** `--feature-mode text` writes LLM output into buffer `feature`
+(format-drift control). Default `scenario_id` for `man_full`; never mix into the primary H0 arm.
+
+## Code entry
+
+```python
+from experiments.man_paper.phase_ii_seed import seed_h0_graph
+
+cortex = seed_h0_graph(llm, store=store, dcn=dcn)
+# cortex.id → bootstrap_id for MAN lifecycle / plasticity
+```