test(v2.1.0): certify regime-conditioned EURUSD triangle reconstruction end to end

[dmidlo]

Outcome

Publish the release-grade scientific and operational dossier for the first online product: historically anchored, regime-conditioned EURUSD/GBPUSD/EURGBP ASCII tick reconstruction under a qualified modern reference regime.

Parent: #430
Audit authority: #459
Depends on: #18, #80, #448, #460-#468

Audit status

The repository contains a certification policy/evaluator and unit tests. No real certification campaign or dossier has passed. Synthetic scalar observations, fixture artifacts, or the existence of an evaluator are not certification evidence.

Broker adaptation is excluded from the v2.1.0 certification. The product must identify its delivery mode as modern reference/unconditioned reference and must not make broker-specific claims. Broker capture, fingerprint, and transfer remain optional separately qualified extensions.

Certification scope

• Instruments: EURUSD, GBPUSD, EURGBP.
• Granularity: ASCII tick only.
• Coverage: source inventory currently shows common monthly support from 2002-03 through 2026-06; re-inventory and hash at execution time.
• Inputs: immutable raw caches plus qualified real artifacts from feat(v2.1.0): fit active-time multivariate feed epochs on real triangle evidence #460-feat(v2.1.0): build and qualify the real modern reference-motif library #464.
• Execution: first-party plan, default worker handlers, public CLI/API from feat(v2.1.0): compose a first-party tick reconstruction plan and artifact graph #465-feat(v2.1.0): expose reconstruction through the CLI and typed API #467.
• Modes: ex-post reconstruction and separately enforced ex-ante-valid views where context support allows them.
• Generator: qualified empirical motif method plus transparent controls.
• Output: ensemble Parquet partitions and compact manifests; no claim of recovered historical truth.

Evidence gates

Source and artifact readiness

• Source inventory, duplicate-dimension audit, hashes, common range, and cache readability reconcile.
• v2 multivariate epoch artifact and stability evidence from feat(v2.1.0): fit active-time multivariate feed epochs on real triangle evidence #460 are present and valid.
• Point-in-time context corpus from feat(v2.1.0): ingest a versioned point-in-time market-context corpus #461 has documented coverage and missingness.
• Point-in-time CFTC positioning corpus from feat(v2.1.0): ingest point-in-time CFTC Commitments of Traders positioning state #468 has documented coverage, availability confidence, and immutable lineage.
• Fitted observation operators from feat(v2.1.0): calibrate historical observation operators from real dense windows #462 are supported and non-identity where a calibrated missing-event claim is made.
• Benchmark corpus/gate version from test(v2.1.0): publish the reverse-degradation benchmark corpus and promotion gates #463 predates promotable candidate results.
• Qualified motif artifact from feat(v2.1.0): build and qualify the real modern reference-motif library #464 passes split/leakage and support checks.

Scientific integrity

• Every recorded historical tick remains byte/logically unchanged in value and identity.
• Every synthetic tick carries deterministic identity, origin, method, confidence, generation ID, neighboring raw anchors, and constraint decisions.
• Point-in-time/leakage audit passes for every ex-ante claim.
• Reverse-degradation hard gates pass on untouched blocked holdouts and negative controls fail as expected.
• Event-count, duration, burst/quiet, update-type, spread, stale-run, timestamp-precision, path, session/event, and epoch-conditioned scorecards pass predeclared tolerances.
• Triangle/inverse/event-synchronization and stale-alignment gates pass without altering observed anchors.
• Ensemble calibration, diversity, refusal, unsupported-region, between-seed, and between-window uncertainty are reported.
• Results are explicitly labeled plausible counterfactual ensembles, never the actual missing ticks.
• Unsupported context or contradictory anchors fail closed.

Product and operations

• A full-range synchronized plan and resource preflight complete through public surfaces.
• Representative sparse, transitional, dense, ordinary, rollover, news-supported, shock, and refusal windows execute through the default worker.
• At least one substantial multi-period run commits queryable final Parquet and compact manifests.
• A mid-run worker/server failure resumes without duplicate or missing committed partitions.
• Cancellation leaves no publishable partial partition.
• Clean replay reproduces logical content hashes across supported concurrency settings.
• Actual peak memory, scratch, runtime, candidate amplification, final rows, and storage remain inside predeclared budgets.
• Corruption, stale artifact, missing context, invalid information mode, quota overflow, and partial triangle tests fail closed.
• Activity/volume semantics (feat(v2.1.0): define synthetic activity and volume-proxy semantics for reconstructed ticks #80) and downstream bars (feat(v2.1.0): derive standard candlestick bars from final reconstruction streams #18) reconcile from final ticks; bars are not reconstruction inputs.
• Strategy-sensitivity results (feat(v2.1.0): evaluate present-day strategy sensitivity on reconstructed history #448) report uncertainty and do not select an automatic winner.
• CLI and API show bounded final tick tables and the complete evidence chain.

Release discipline

• Human-readable methodology, nonclaim, limitations, and machine-readable evidence manifest are published.
• The environment installs every declared test dependency; no suite is silently skipped due to missing packages.
• Full pre-commit hooks and all test suites pass.
• Coverage runs once at the dev-to-main promotion boundary under repository policy.
• TestPyPI publication occurs from dev and passes the local simple-registry installation/import/CLI smoke preflight.
• Only after all gates pass is dev promoted to main and the same release artifact published to PyPI.

Closure rule

This issue cannot close because the pipeline completes, output looks plausible, unit tests pass, or an accepted-limitation field is populated. Every gate must link to reproducible real evidence. A limitation may be accepted only when it narrows the product claim and does not contradict immutable anchors, information safety, benchmark validity, or operational correctness.

SemVer

Release acceptance issue for v2.1.0.