research(v2.4.0): implement a constrained Schrödinger-bridge proposal engine

[dmidlo]

Outcome

Explore path-space transport between historical constraints and a broker-conditioned modern delivery distribution as an opt-in research challenger after the primary product path has been certified.

Parent: #430
Depends on: #431, #433, #436, #441, #445, #449 (all closed before implementation)

Falsifiable hypothesis

On untouched reverse-degradation validation and final holdouts, a bounded discrete Schrödinger bridge will improve joint timing/mark/path/triangle diagnostics over dense/no-fill, linear interpolation, empirical motif, and simpler point-process comparators without violating immutable anchors, forbidden intervals, information boundaries, or resource limits.

Failure to converge, missing endpoint support, stale broker/epoch context, or a hard-constraint violation is a refusal—not a generated path. A visually plausible output is not evidence.

Fixed prototype boundary

Implement a finite-state Markov Schrödinger bridge over bounded event windows:

• state atoms are normalized time bin × destination symbol × quote-transition mark;
• the reference path law is a versioned, train-only, first-order Markov kernel;
• source and broker-conditioned endpoint marginals are joined by entropic iterative proportional fitting/Sinkhorn scaling;
• sampled intermediate states use the conditional Markov-bridge law, not marginal interpolation;
• cardinality is a separately versioned bounded count law;
• broker profile selection and transfer strength define the target marginal explicitly;
• EURUSD/GBPUSD/EURGBP currency-exposure distance contributes a separate transport-cost term and final triangle validation remains authoritative;
• observed anchors stay external and byte/logically immutable, while quarantine/closed intervals admit no synthetic event;
• whole-window solving is the scientific reference; streaming carries only bounded strict-prior state and reports boundary approximation error.

The implementation must remain dependency-free/CPU-bounded unless a later issue authorizes another runtime.

Required evidence

• content-addressed config, broker-target, dataset, solver/checkpoint, fit, generation, and lineage contracts;
• time-ordered train/tune handling with row-free validation/final protected manifests and leakage/overlap checks;
• convergence trace, marginal residuals, endpoint support, regularization/transport cost, numerical ranges, approximation error, wall time, peak memory, work, and refusal reason;
• deterministic generation for fixed sources/config/member and uncertainty across independent members;
• shared benchmark and historical-carving integration without a private constraint path;
• a retained real-corpus reverse-degradation comparison, or an explicit evidence-backed refusal if the predeclared solver/support/resource gates cannot be met.

Acceptance criteria

• A falsifiable research hypothesis and transparent simpler comparators are predeclared.
• Observed anchors and forbidden intervals are exact constraints.
• Cross-currency and broker conditioning have explicit mathematical roles.
• Solver convergence, approximation error, numerical stability, and resource use are reported.
• Streaming/window-boundary approximations are documented and tested.
• Reverse-degradation holdouts test more than marginal distribution matching.
• Non-convergence/off-support cases refuse rather than emit plausible-looking paths.
• The prototype cannot become a production default without a separate evidence-backed promotion issue.

Non-goals

This issue does not replace the certified empirical method, recover actual missing ticks, prove broker adaptation generally, introduce a Gaussian/neural diffusion runtime, or authorize production promotion.

SemVer

Optional minor research feature in the v2.4.0 challenger train. No standalone package release is required; any later release follows the repository's dev → TestPyPI → main → PyPI policy.

[dmidlo]

Implemented and validated on dev in ee21453.

Closure evidence:

• Added the dependency-free finite-state Markov/Sinkhorn Schrödinger-bridge challenger, strict versioned/content-addressed contracts, broker-target and cross-currency conditioning, bounded exact refusal behavior, deterministic generation, strict-prior streaming evidence, and shared benchmark/carving integration.
• Added direct unit coverage for immutable anchors, exact forbidden intervals, non-convergence/off-support refusal, protected holdouts, tamper detection, triangle projection, deterministic/independent members, bounded history, and shared carving.
• Retained and documented a real-histdata reverse-degradation campaign: 12 held-out windows, 2 members, 552 bridge metrics, 0 execution failures/refusals/anchor violations/unsupported emissions.
• The predeclared joint hypothesis was falsified. Event-count, path-variation, triangle-residual, and update-transition gates failed; the report is provisional, promotion_eligible=false, and automatic_winner=false. No production promotion is authorized.
• Candidate report: reverse-degradation-candidate-report:sha256:3490191d4a423bfb95ee373a9bf579f66f94e21cf6c52bb6e9440fe3007d6fc6
• Full pre-commit battery passed. Full pytest: 2143 passed, 4 skipped, 0 failed in 33:04.

All eight acceptance criteria are satisfied.