[SignalLayerLabs]

Problem and evidence

Curriculum drawdown currently covers closed trades, so a later profitable exit can conceal an adverse open-position path. The price-only hold reference excludes execution costs and assumes full investment.

Change

• Add report-only, versioned bar-close NAV/drawdown from existing frozen fills, with cash reserved and fees booked once at the recorded clocks.
• Compare policy and fixed expert with a delayed-opening hold using the same capital, window, commissions and capacity model. Preserve residual cash after partial entry; keep mark NAV distinct from estimated net liquidation.
• Withhold values for invalid/missing entry or mark paths and unverified exit capacity. Compute full-resolution statistics before curve compaction; ignore future prices/exits in prefix values and hashes.
• Store reports in the worker, including a one-time evaluation refresh for exhausted packs without relearning or checkpoint changes. Show optional, backward-compatible Research tables, per-pair quote units and current-episode-only prospective valuation.
• Document the pinned LEAN analysis and independent Python implementation; no LEAN/.NET runtime or copied source/data dependency.

Verification

• Final backend/migration suite: 1,037 passed in 121.47s; existing Pydantic/UTC deprecation warnings remain.
• Frontend: 44 tests passed; TypeScript and static production build passed (63 pages).
• Independent review: two report-only P2 findings reproduced RED, fixed and regression-tested; no remaining P1/P2 findings.
• Fresh durable-worker replay on repository SPY/QQQ and local hashed Yahoo Chart EURUSD/GBPUSD inputs: 32 cycles, 1,109 dependent expert outcomes, zero duplicate updates after a new DB session. All source hashes, model checkpoints, learning counts and original policy/fixed-expert metrics exactly match the preceding source-recovery receipt.
• QQQ policy +0.345%, bar-close drawdown 1.110%, versus modeled hold net-liquidation +66.472%; SPY policy −0.345%, drawdown 0.345%, versus hold +42.131%. Both USA policies underperform hold. FX full-path values are WITHHELD for preserved invalid bars; confirmation remains unopened. See docs/evidence/lean-nav-evaluation.json.
• Space metadata, packaged-source audit and diff whitespace checks passed.

Risk and rollback

This improves measurement, not model skill or demonstrated alpha. Costs/liquidity are modeled; hold is a cash-notional opening scenario and terminal sale is an estimate. Bar-close drawdown excludes intrabar losses. No dividends/carry/taxes or cross-currency aggregation. Revert this commit to remove optional reports/UI; immutable source, outcome and checkpoint identities remain intact. No schema migration.

Production PR15 is deployed at 8789d12, but useful curriculum learning is not yet verified: an observed scheduled event deferred before entering the worker for resource_pressure, while its outer status was mislabeled completed. This PR preserves that guard; runtime resource diagnosis is separate and remains open.

• No credentials or restricted data are included.
• GET endpoints remain read-only.
• No frontend render triggers heavy computation.
• Financial claims are evidence-bound.
• Tests cover the changed behavior.