feat: establish ESDP-light contracts and feasibility evidence

[jimmlucas]

What changed

• establishes cost, provenance, and temporal-causality contracts for ESDP-light
• adds deterministic esdp light-observe and esdp light-history commands
• extracts reference-free FASTA metrics and records artifact SHA-256 identities
• rejects incomplete, duplicated, skipped, non-finite, and out-of-domain histories
• packages the light modules in the Python distribution and production image
• adds development-only grouped feasibility, formulation, endpoint, and measurement-validity studies with complete out-of-fold artifacts

Why

ESDP-light aims to decide whether another polishing round is useful without reproducing expensive BUSCO/QV analysis during inference. That requires explicit metric provenance, causal temporal features, a defensible stopping endpoint, and validation grouped by biological sample.

Observation and history generation are operational. Adaptive stopping remains deliberately unavailable until the scientific release gates are satisfied.

Scientific evidence

All studies use only the 32 frozen development samples (127 complete R1-R5 trajectories). The nine-sample test set remains sealed.

Three-class feasibility

At the predeclared 0.60 threshold, light core achieved 0.345 balanced accuracy, 7.9% unsafe stopping, and 0.291 rounds saved. This rejected a release model and motivated a direct sequential formulation.

Direct binary stopping

Predicting CONTINUE versus STOP_ELIGIBLE significantly reduced light-core absolute round error by 0.157 rounds. Balanced accuracy remained 0.406. A historical leaky control reached 0.844 but fell to 0.540 after causal reconstruction, demonstrating substantial temporal-leakage inflation.

Future-material-benefit endpoint

The endpoint assigns R1/R2/R3/R4/R5 to 88/17/10/7/5 trajectories, compared with the legacy 3/20/22/11/71 distribution. Light core reaches 0.502 balanced accuracy and per-round ROC AUC of 0.685-0.761, showing real low-cost cross-sample signal. At threshold 0.70 it still has 14.2% unsafe stops and 12.6% QV/BUSCO failures while saving 3.024 rounds.

Measurement-validity gate

The audit finds that qv = -10*log10(error_rate) exactly and assembly_error = abs(assembly_frac - 1); these pairs are redundant outcomes. All development trajectories use 116 BUSCO markers, so the >1-point threshold requires at least two markers. However, 12/127 trajectories show only a transient >1-point BUSCO gain, no technical repeatability can be estimated, and BUSCO version/lineage/date/mode/options are absent from the table.

The versioned materiality contract therefore blocks model retraining, held-out testing, light-decide, and adaptive nf-core/bacass stopping. Non-decision observation/history integration remains allowed.

Scientific boundaries

• no release model artifact or production stopping command is created
• experimental models are development-only
• v1 model, manuscript outputs, and published behavior remain unchanged
• expensive retrospective quality metrics cannot become ESDP-light inference features
• provisional alignment metrics remain excluded from generated histories
• threshold sweeps describe safety-savings frontiers and are not selection procedures
• the held-out test remains sealed until measurement, endpoint, calibration, and safety contracts are frozen

Validation

• full suite: 192 passed
• frozen split: 32 development and 9 held-out biological samples
• held-out samples used by every experiment: 0
• complete grouped out-of-fold predictions at every decision round
• strict JSON reports include dataset/split checksums without NaN
• temporal prefix invariance prevents future rounds changing earlier features
• Python wheel and production Docker image include all ESDP-light runtime modules
• light-observe R1/R2 -> light-history executes successfully in Docker

The 30 full-suite warnings are existing FastAPI, Pydantic, and httpx deprecations unrelated to ESDP-light. This PR remains a draft pending a validated measurement/materiality contract and an acceptable asymmetric premature-stop safety gate.