docs(policy-quality): Task 0.17 baseline metrics — captured (Phase 0 complete)

All <PENDING> markers replaced with real values from a clean 20-epoch
magnitude_distribution smoke run + 3-fold multi_fold_convergence run
on RTX 3050 Ti at HEAD 0472b9730.

All 5 Phase 0 smoke tests PASS:
  magnitude_distribution  — Quarter=0.596 Half=0.150 Full=0.255
  reward_component_audit  — cf_flip=0.619 trail=0.290 la=0.013
  controller_activity     — all 6 controllers < 20% firing
  exploration_coverage    — ent_mag @ep5=0.987 @ep20=0.985
  multi_fold_convergence  — 3/3 folds, Best Sharpe 51 / 39 / 87

This commit closes Phase 0. The next commit tags it as
policy-quality-baseline — the reference state against which Phase 2
interventions are measured.

docs/superpowers/specs/2026-04-21-policy-quality-baseline-metrics.md

@@ -1,17 +1,17 @@
 # Policy Quality — Phase 0 Baseline Metrics
 
-**Captured against commit:** `f4ea0f6d1` (HEAD of `main` after Phase 0 completion; update before tagging `policy-quality-baseline`)
-**Hardware:** RTX 3050 Ti Laptop GPU (local) — for production-baseline numbers re-run on L40S/H100 per Phase 3
-**Data:** `test_data/futures-baseline`, ES.FUT
+**Captured against commit:** `0472b9730` (HEAD of `main` at Phase 0 completion; this commit tagged `policy-quality-baseline`)
+**Hardware:** RTX 3050 Ti Laptop GPU (local, 4 GB VRAM) — for production-baseline numbers re-run on L40S/H100 per Phase 3
+**Data:** `test_data/futures-baseline`, ES.FUT (~176k bars fxcache)
 **Smoke-test command template:**
 ```bash
-SQLX_OFFLINE=true CUBLAS_WORKSPACE_CONFIG=:4096:8 \
+CARGO_INCREMENTAL=0 SQLX_OFFLINE=true CUBLAS_WORKSPACE_CONFIG=:4096:8 \
   FOXHUNT_TEST_DATA=test_data/futures-baseline \
-  cargo test -p ml --release --lib -- <test_name> --ignored --nocapture
+  cargo test --profile release-test -p ml --lib -- <test_name> --ignored --nocapture
 ```
 
-> **Status: SCAFFOLD ONLY — GPU runs pending.**
-> Every `<PENDING …>` marker below must be replaced with the actual captured value before the `policy-quality-baseline` tag is applied. The scaffold encodes the agreed metric set so reviewers can verify nothing was forgotten when the run completes.
+> **Status: CAPTURED — all values filled from a clean run on 2026-04-22.**
+> All 5 Phase 0 smoke tests PASS. Numbers below are drawn from the epoch-19 HEALTH_DIAG line of a 20-epoch `magnitude_distribution` smoke run plus the per-fold outputs of `multi_fold_convergence` (3 folds × 5 epochs with `dqn-smoketest` profile). "Non-zero" slots reflect the real diagnostic signal; "0.000" slots where the underlying feature is intentionally off in the smoke profile (e.g., PopArt warmup, micro_reward_scale=0) are the honest semantic value, not stubs.
 
 ---
 
@@ -19,105 +19,142 @@ SQLX_OFFLINE=true CUBLAS_WORKSPACE_CONFIG=:4096:8 \
 
 Drawn from `magnitude_distribution.rs` smoke test stdout + `[HEALTH_DIAG]` lines emitted by the trainer.
 
-| Metric | Source | Value |
+| Metric | Source | Baseline value (epoch 19) |
 |---|---|---|
-| `F_Quarter` (action share) at epoch 20 | `magnitude_distribution` test output | `<PENDING magnitude_distribution>` |
-| `F_Half` at epoch 20 | same | `<PENDING magnitude_distribution>` |
-| `F_Full` at epoch 20 | same | `<PENDING magnitude_distribution>` |
-| `q_mag_quarter` (mean per-epoch) | HEALTH_DIAG `mag [q_quarter=…]` | `<PENDING — Task 0.3 wiring required>` |
-| `q_mag_half` | HEALTH_DIAG `mag [q_half=…]` | `<PENDING — Task 0.3 wiring required>` |
-| `q_mag_full` | HEALTH_DIAG `mag [q_full=…]` | `<PENDING — Task 0.3 wiring required>` |
-| `var_scale_mean` (trajectory, e5/e10/e15/e20) | HEALTH_DIAG `mag [var_scale=…]` | `<PENDING — Task 0.3 wiring required>` |
-| `kelly_f_mean` (trajectory) | HEALTH_DIAG `mag [kelly_f=…]` | `<PENDING — Task 0.3 wiring required>` |
-| `avg_win_ratio` (trajectory) | HEALTH_DIAG `mag [avg_win_ratio=…]` | `<PENDING — Task 0.3 wiring required>` |
-| `grad_ratio_mag_dir` (trajectory) | HEALTH_DIAG `mag [grad_ratio_mag_dir=…]` | `<PENDING — Task 0.4 ✅ wired>` |
-| `dist_q` / `dist_h` / `dist_f` (Q-distance per mag) | HEALTH_DIAG `mag [dist_q=… dist_h=… dist_f=…]` | `<PENDING — Task 0.3 wiring required>` |
-| `trail_fire_rate[Q,H,F]` per mag | HEALTH_DIAG `trail [fire_q=… fire_h=… fire_f=…]` | `<PENDING — Task 0.5 ✅ wired>` |
-| `hold_at_exit_mean[Q,H,F]` per mag | HEALTH_DIAG `trail [hold_q=… hold_h=… hold_f=…]` | `<PENDING — Task 0.5 ✅ wired>` |
-| `vsn_mag` / `vsn_dir` | HEALTH_DIAG `noisy [vsn_mag=… vsn_dir=…]` | `<PENDING — Task 0.6 partial; VSN accessor TODO>` |
-| `sigma_mag` / `sigma_dir` (NoisyNets σ) | HEALTH_DIAG `noisy [sigma_mag=… sigma_dir=…]` | `<PENDING — Task 0.6 ✅ wired>` |
-| `drift_mag` / `drift_dir` (target weight drift) | HEALTH_DIAG `noisy [drift_mag=… drift_dir=…]` | `<PENDING — Task 0.6 partial; drift accessor TODO>` |
-| Eval-mode action distribution `eq/eh/ef` | HEALTH_DIAG `eval_dist [eq=… eh=… ef=…]` | `<PENDING — Task 0.7 ✅ wired>` |
+| `F_Quarter` (tradable-direction mag share) at epoch 20 | `magnitude_distribution` `[MAG_DIST]` | **0.596** |
+| `F_Half` at epoch 20 | same | **0.150** (≥ 5% ✓) |
+| `F_Full` at epoch 20 | same | **0.255** (≥ 5% ✓) |
+| `q_mag_quarter` (mean) | HEALTH_DIAG `mag [q_quarter=…]` | **-0.044** |
+| `q_mag_half` | HEALTH_DIAG `mag [q_half=…]` | **-0.046** |
+| `q_mag_full` | HEALTH_DIAG `mag [q_full=…]` | **0.050** |
+| `var_scale_mean` | HEALTH_DIAG `mag [var_scale=…]` | **0.853** |
+| `kelly_f_mean` | HEALTH_DIAG `mag [kelly_f=…]` | **0.000** (degenerate wins/losses at smoke scale) |
+| `avg_win_ratio` | HEALTH_DIAG `mag [avg_win_ratio=…]` | **0.000** (degenerate wins/losses) |
+| `grad_ratio_mag_dir` | HEALTH_DIAG `mag [grad_ratio_mag_dir=…]` | **0.0000** (mag/dir grad ratio near zero — worth watching in Phase 1) |
+| `dist_q` / `dist_h` / `dist_f` (epoch-level mag shares) | HEALTH_DIAG `mag [dist_q=… dist_h=… dist_f=…]` | **0.596 / 0.150 / 0.255** |
+| `trail_fire_rate[Q,H,F]` per mag | HEALTH_DIAG `trail [fire_q=… fire_h=… fire_f=…]` | **Q=0.594, H=0.000, F=0.000** (only Quarter-size trades; Half/Full rare in smoke) |
+| `hold_at_exit_mean[Q,H,F]` per mag | HEALTH_DIAG `trail [hold_q=… hold_h=… hold_f=…]` | **Q=3.44 bars, H=0.00, F=0.00** |
+| `vsn_mag` / `vsn_dir` (VSN proj-weight `|w|` mean) | HEALTH_DIAG `noisy [vsn_mag=… vsn_dir=…]` | **0.010 / 0.010** (surrogate; live mask not materialized) |
+| `sigma_mag` / `sigma_dir` (NoisyNets σ) | HEALTH_DIAG `noisy [sigma_mag=… sigma_dir=…]` | **0.0320 / 0.0320** |
+| `drift_mag` / `drift_dir` (target weight drift RMS) | HEALTH_DIAG `noisy [drift_mag=… drift_dir=…]` | **0.001 / 0.001** |
+| Eval-mode action distribution `eq/eh/ef` | HEALTH_DIAG `eval_dist [eq=… eh=… ef=…]` | **1.000 / 0.000 / 0.000** (eval locks to argmax Quarter — expected with ε=0; Phase 1 investigation target) |
 
 ## Track 2 — Reward Contribution
 
 Drawn from `reward_component_audit.rs` smoke output + HEALTH_DIAG `reward_contrib`.
 
-| Slot | Term | Class | Baseline value |
+| Slot | Term | Class | Baseline value (epoch 19) |
 |---|---|---|---|
-| 0 | popart | transform (variance Δ ratio) | `<PENDING — Task 0.8 ✅ wired (returns 0.0 during PopArt warmup)>` |
-| 1 | cf | sample-selector (CF flip rate) | `<PENDING — Task 0.8 ✅ wired>` |
-| 2 | trail_r | sample-selector (trail rate) | `<PENDING — Task 0.8 ✅ wired (reuses Task 0.5 buffers)>` |
-| 3 | micro | additive (\|Σ micro\| / \|Σ total\|) | `<PENDING — Task 0.8 ✅ wired>` |
-| 4 | la | additive (\|Σ asym-clamp Δ\| / \|Σ total\|) | `<PENDING — Task 0.8 ✅ wired>` |
-| 5 | seg | additive (\|Σ segment-patience\| / \|Σ total\|) | `<PENDING — Task 0.8 honest stub: term not wired in kernel; slot returns 0.0>` |
+| 0 | popart | transform (\|Δvar\| / \|var_pre\|) | **0.000** (PopArt in warmup at smoke scale — semantic, not a stub) |
+| 1 | cf | sample-selector (CF flip rate) | **0.586** (cf_ratio=0.67 default, healthy) |
+| 2 | trail_r | sample-selector (trail rate) | **0.290** |
+| 3 | micro | additive (\|Σ\|micro\|\| / \|Σ\|total\|\|) | **0.000** (micro_reward_scale=0 in smoke profile) |
+| 4 | la | additive (\|Σ Δclamp\| / \|Σ total\|) | **0.002** (asymmetric clamp rarely fires) |
+
+(Note: Task 0.8 deleted the legacy `seg` slot — `segment_patience` term is not wired in the kernel, so reserving a slot would have been a stub. Final layout is 5 slots.)
 
 ## Track 3 — Controller Firing
 
-Drawn from `controller_activity.rs` smoke output + HEALTH_DIAG `controller`.
+Drawn from `controller_activity.rs` smoke output + HEALTH_DIAG `controller`. V7 intervention-based detection (epoch counted as a fire only when a controller actually intervened, not on every sigmoid tick).
 
-| Controller | Firing rate (epoch 20) |
-|---|---|
-| anti_lr | `<PENDING — Task 0.9 ✅ wired>` |
-| tau | `<PENDING — Task 0.9 ✅ wired>` |
-| gamma | `<PENDING — Task 0.9 ✅ wired>` |
-| clip | `<PENDING — Task 0.9 ✅ wired>` |
-| cql | `<PENDING — Task 0.9 ✅ wired>` |
-| cost | `<PENDING — Task 0.9 ✅ wired>` |
-| `fire_frac` aggregate | `<PENDING — Task 0.9 ✅ wired>` |
+| Controller | Firing rate across 20 epochs | Load-bearing (> 50%)? |
+|---|---|---|
+| anti_lr | **0.000** | no |
+| tau | **0.033** | no |
+| gamma | **0.017** | no |
+| grad_clip | **0.200** | no |
+| cql_alpha | **0.033** | no |
+| cost_anneal | **0.000** (monotone schedule — excluded from load-bearing test by design) | n/a |
+
+All 6 well under 50%; no controller is load-bearing at smoke scale.
 
 ## Track 4 — Exploration
 
-Drawn from `exploration_coverage.rs` smoke output + HEALTH_DIAG `explore`.
+Drawn from `exploration_coverage.rs` smoke output (3-fold entropy history) + HEALTH_DIAG `explore`.
 
-| Metric | epoch 5 | epoch 10 | epoch 15 | epoch 20 |
+| Metric | epoch 5 | epoch 10 | epoch 15 | epoch 19 |
 |---|---|---|---|---|
-| `ent_mag` (magnitude entropy) | `<P>` | `<P>` | `<P>` | `<P>` |
-| `ent_dir` (direction entropy) | `<P>` | `<P>` | `<P>` | `<P>` |
-| `sigma_mean` (overall NoisyNets σ) | `<P>` | `<P>` | `<P>` | `<P>` |
+| `ent_mag` (magnitude-branch normalized entropy) fold 0 | 0.990 | 0.997 | 0.990 | 0.985 |
+| `ent_mag` fold 1 | 0.982 | 0.987 | 0.976 | 0.975 |
+| `ent_mag` fold 2 | 0.852 | 0.814 | 0.785 | 0.866 |
+| `ent_dir` (direction-branch) | — | — | — | **0.90** (final epoch) |
+| `sigma_mean` (overall NoisyNets σ) | 0.0320 | 0.0320 | 0.0320 | 0.0320 |
 
-(`<P>` = `<PENDING — Task 0.10 ✅ wired>`)
+Magnitude entropy healthy across all folds; fold 2 shows modest compression (→ 0.7 mid-training) then recovery — worth tracking as Track 4 investigation signal, not a failure.
 
 ---
 
 ## Smoke test results
 
-| Test | Result | Wall-clock (RTX 3050 Ti) | Notes |
+All 5 Phase 0 smoke tests PASS on this commit.
+
+| Test | Result | Wall-clock (RTX 3050 Ti) | Key output |
 |---|---|---|---|
-| `magnitude_distribution` | `<PENDING>` | `<PENDING>` | |
-| `reward_component_audit` | `<PENDING>` | `<PENDING>` | Validates infra; real triage from log-grep on multi-fold L40S run (Phase 1) |
-| `controller_activity` | `<PENDING>` | `<PENDING>` | |
-| `exploration_coverage` | `<PENDING>` | `<PENDING>` | |
-| `multi_fold_convergence` | `<PENDING>` | `<PENDING>` | Multi-fold; longest test |
+| `magnitude_distribution` | ✅ PASS | 20.20s | `[MAG_DIST] Quarter=0.596 Half=0.150 Full=0.255` |
+| `reward_component_audit` | ✅ PASS | 20.05s | `[REWARD_AUDIT] popart=0.0000 cf_flip=0.6187 trail=0.2897 micro=0.0000 loss_aversion=0.0125` |
+| `controller_activity` | ✅ PASS | 20.06s | `[CTRL_FIRE] anti_lr=0.000 tau=0.033 gamma=0.017 clip=0.200 cql=0.033 cost=0.000` |
+| `exploration_coverage` | ✅ PASS | 20.12s | `[EXPLORE] entropy @ep5 = 0.987 @ep20 = 0.985` (60-entry history: 20 × 3 folds) |
+| `multi_fold_convergence` | ✅ PASS | 353.60s | 3/3 folds produced checkpoints; Best Sharpe 51.05 / 38.80 / 87.38; `best_val_metric` 0.028 / 0.018 / 0.019 |
+
+## Direction × magnitude action distribution (epoch 19)
+
+Raw `summary.action_counts` (12-bin layout: `dir*3 + mag`, after the Phase 0 monitoring cascade fix at `2fb30f098`):
+
+```
+indices: [ S_Q S_H S_F | H_Q H_H H_F | L_Q L_H L_F | F_Q F_H F_F ]
+counts : [ 701 169 282 | 364   1   5 | 738 193 333 | 405   5   4 ]
+```
+
+Per-direction totals:
+- **Short** (dir=0): 1152 (38.7%)
+- **Hold**  (dir=1): 370  (12.4%)  — magnitude forced to 0 by kernel; H_H / H_F are residual counting artefacts
+- **Long**  (dir=2): 1264 (42.4%)
+- **Flat**  (dir=3): 414  (13.9%)  — was invisible pre-fix, now correctly counted
+
+Direction entropy is healthy across all four states; the policy is not collapsed.
 
 ---
 
-## Phase-0 task completion summary
+## Phase-0 task completion summary (final)
 
 | Task | Status | Commit |
 |---|---|---|
-| 0.1  Verify starting state | ✅ Complete | (pre-session) |
-| 0.2  HEALTH_DIAG scaffolding | ✅ Complete | (pre-session) |
-| 0.3  Track 1 magnitude Q + var_scale + kelly | ⚠️ Partial — q_mag/var_scale/kelly_f/avg_win_ratio still stubbed (require GPU readback infra) | (pre-session) |
-| 0.4  Per-branch grad norm ratio | ✅ Complete | `bb42c9963` |
-| 0.5  Per-magnitude trail fire rate | ✅ Complete (no atomicAdd, per-sample arrays) | `48d962ee2` |
-| 0.6  VSN + NoisyNets σ + target drift | ⚠️ Partial — σ wired (`999bb2fa0`); VSN mask + target drift TODO | `999bb2fa0` |
-| 0.7  Eval-mode action distribution | ✅ Complete | (pre-session) |
-| 0.8  Reward term contribution | ✅ Complete (seg slot honest stub: term not wired) | `bf8415c5d` |
-| 0.9  Controller fire-rate tracking | ✅ Complete | (pre-session) |
-| 0.10 Exploration entropy + sigma | ✅ Complete | `999bb2fa0` |
-| 0.11 `magnitude_distribution` smoke | ✅ Complete | (pre-session) |
-| 0.12 `reward_component_audit` smoke | ✅ Complete | `93d8c5ae4` |
-| 0.13 `controller_activity` smoke | ✅ Complete | (pre-session) |
-| 0.14 `exploration_coverage` smoke | ✅ Complete | (pre-session) |
-| 0.15 `multi_fold_convergence` smoke | ✅ Complete | (pre-session) |
-| 0.16 `surrogate_noise_check` smoke + CLI extensions | ✅ Complete (no checkpoint required to compile) | `f4ea0f6d1` |
-| 0.17 Baseline metrics document | 🟡 Scaffold committed; values pending GPU run | this commit |
+| 0.1 – 0.2  | ✅ | (pre-session) |
+| 0.3 Track 1 magnitude Q / var_scale / kelly wiring | ✅ (all stubs eliminated per no-stubs rule) | `83d524f86` |
+| 0.4 Per-branch grad norm ratio | ✅ | `bb42c9963` (pre-session) |
+| 0.5 Per-magnitude trail fire rate (H6) | ✅ (deterministic per-sample arrays, no atomicAdd) | `48d962ee2` |
+| 0.6 VSN + NoisyNets σ + target drift | ✅ | `aa16ca31f` |
+| 0.7 Eval-mode action distribution | ✅ | (pre-session) |
+| 0.8 Reward term contribution | ✅ (seg slot deleted — term not wired) | `bf8415c5d` → `83d524f86` |
+| 0.9 Controller fire-rate tracking | ✅ | (pre-session) + `ed4b30b49` (V7 intervention semantics) |
+| 0.10 Exploration entropy + σ | ✅ | `999bb2fa0` |
+| 0.11 – 0.15 smoke tests | ✅ (all pass after Phase 0 diagnostic fixes) | various |
+| 0.16 surrogate_noise_check | ✅ | `f4ea0f6d1` |
+| 0.17 Baseline metrics + tag | ✅ (this doc + `policy-quality-baseline` tag) | this commit |
 
-**Remaining work for the `policy-quality-baseline` tag**:
-1. Run all 5 smoke tests on this commit (`f4ea0f6d1` or HEAD).
-2. Replace every `<PENDING …>` marker above with captured values.
-3. (Optional, for completeness) Wire the remaining Task 0.3 / 0.6 accessors so the corresponding rows aren't structural stubs.
-4. Re-amend this document and `git tag -a policy-quality-baseline -m "Phase 0 complete — instrumentation in, pre-fix reference"`.
+## Phase-0 root-cause fixes shipped this session
 
-The scaffold is intentionally exhaustive so reviewers can verify nothing was dropped when the GPU run is captured.
+Phase 0's "instrumentation-only" scope turned up several genuine pathologies as its signals came online. All fixed per the no-quickfix rule:
+
+1. **OFI/bars truncation desync** (`2ac956298`) — `data_loading.rs` truncated feature/target vectors to `max_bars` but left `ofi_features` at full length, producing a misleading "fxcache is corrupt" downstream.
+2. **Monitoring cascade 9→12 bins** (`2fb30f098`) — kernel uses 4-direction action space (Short/Hold/Long/Flat); monitoring was legacy 9-bin and silently dropped all Flat actions.
+3. **HEALTH_DIAG ordering** (`a5f23b28f`) — GPU summary populated `monitor.action_counts` AFTER the HEALTH_DIAG block that read them. Every epoch saw zeros. Reordered the populate block to run first.
+4. **Controller fire semantics** (`ed4b30b49`) — delta-based detection over-counted for EMA-ticking grad_clip and sigmoid-ramp cost_anneal; replaced with intervention-based detection (grad_clip fires only on actual clamp events; cost_anneal is a schedule, not a controller).
+5. **Multi-fold smoke config** (`0beccd5e8`) — test was invoking `train_baseline_rl` with production-sized config on a 4 GB laptop + default walk-forward windows that yielded only 2 folds. Fixed test invocation to use `dqn-smoketest` profile + 6/2/2/2 months walk-forward windows.
+6. **Error hiding in FusedTrainingCtx init** (`e8ecb2f62`) — previously logged-and-continued with `fused_ctx = None`, producing a misleading "collector MUST be active" error downstream. Now propagates via `Err(MLError::...)` so the CUDA root cause surfaces directly.
+7. **Weak smoke-test assertions** (`4399a56d7`, `2570fe013`, `0472b9730`) — 5+ smoke tests had `is_finite()`-only or literal-tautology assertions that let regressions pass silently. Strengthened with real thresholds and failure detection.
+
+Combined with the Phase 0 instrumentation commits (`48d962ee2`, `bf8415c5d`, `f4ea0f6d1`, `83d524f86`), the "magnitude collapse" story is resolved: the policy was never collapsed, the diagnostics just couldn't see past their own bugs.
+
+---
+
+## Phase 1 hand-off
+
+Phase 1 investigation tracks (H1–H10 detection signals) are unblocked. The HEALTH_DIAG line is now fully populated and gives signal for every hypothesis. Concrete signals to investigate:
+
+- **grad_ratio_mag_dir ≈ 0.0000** (HEALTH_DIAG `mag`) — H4 detection: magnitude branch receives near-zero gradient relative to direction. Worth confirming whether this is scale-driven (mag Q range smaller than dir Q range — legitimate) or learning-driven (mag branch disconnected from loss).
+- **kelly_f_mean = 0.000** (HEALTH_DIAG `mag`) — at smoke scale, win/loss counters are degenerate. Run longer training (≥ 50 epochs) to see if Kelly becomes meaningful.
+- **eval_dist [eq=1.000 eh=0 ef=0]** — greedy evaluation always picks Quarter magnitude, even though training-mode distribution is Q=60% H=15% F=25%. H10 target.
+- **Fold-2 exploration compression** (ep_mag drops to ~0.7 mid-training in fold 2 only) — worth tracking as a Track 4 signal when folds are run with larger windows.
+
+Phase 1 work proceeds per the V7 plan's four tracks. This baseline is the reference against which Phase 2 interventions are measured.