feat(tools): add external dataset replay harness

[jscott3201]

Summary

• add an optional offline external-dataset inspection and replay harness
• add the first inventory-backed adapter contract for the LBNL simulated PFPU/SFPU dataset
• add a bounded JSON trace mode to the existing Rust verifier so Python never reimplements CXF/CDL logic
• add a stable result schema, manual-acquisition and mapping documentation, and future-adapter guidance
• add only original tiny PFPU/SFPU fixtures; no source archive, inventory PDF, or derived large replay is committed

Architecture and evidence boundaries

Decision	Result
Rule execution	Temporary gated vectors call the existing pinned open-control engine through cxf-verify --trace-json.
Mapping	A required local manifest records case/point/gate inventory evidence, exact columns, units, proxies, dependencies, subtype, labels, fault start, and source version; names are never guessed.
Independence	Command/proof, actual/reference, actual/expected, measurement/setpoint, inlet/outlet, and gate/outcome circularity is rejected. A documented topology-equivalent SFPU primary/fan-flow alias is the only narrow exception.
Baselines	Expected fan flow and coil delta-T require model readiness/version/input/known-good/error/update metadata.
Timeline	Duplicate, out-of-order, and irregular timestamps block replay; gaps and missing required inputs reset both engine state and startup lead. Booleans are never interpolated.
Gating	Six required host-gate categories retain fixed minimum applicability; additional independent inventory-backed gates remain supported.
Metrics	Healthy rates use fault-free evaluable samples. TPR uses declared inventory rule-case pairs, post-fault evaluability, and post-fault alarm onsets; pre-alarmed windows do not count as detections.
Identity	Results record dataset pre/post fingerprint, library revision/dirty state, engine pin and clean build revision, verifier/tool/rule/card/point hashes, and engine-exported content IDs.
Scale/safety	Cases run sequentially; trace batches, sample count, runtime, and output size are bounded. Fallback Cargo execution is offline. Writes are restricted away from source/canonical trees.

Source policy

Source	Use in this PR	Not claimed
Official LBNL simulated FPU dataset, DOI 10.25984/1881324	Adapter target, topology/fault-class inventory contract, future empirical replay	No source archive was locally available; no measured FPR/TPR or threshold validation
Existing FPB rule cards/graphs	Committed rule semantics, content IDs, engine replay targets	No Python reimplementation or threshold change
Original tiny fixtures	Contract, failure-mode, and pinned-engine smoke tests	Not redistributed LBNL data and not field-performance evidence

Validation

• dataset harness contract tests: 32 passed
• Rust trace-boundary tests: 2 passed
• real pinned-engine tiny fixture replay: six FPB rules across one healthy PFPU and one restricted-flow SFPU
• fixture result: zero healthy alarm samples; FPB-0004 detected after 900 s; declared/evaluable/detected target-pair counts all pinned
• result JSON validated against tools/dataset_harness/result.schema.json
• full verifier: 134 rule packages, all scenarios passed
• registry lint: 134 rules OK
• SVG fitment: 135 SVGs clean
• existing simharness: 6 passed
• Cargo test/check and Python compile checks
• mdBook: 134 fault pages, 21 playbooks, 13 point dictionaries
• workflow YAML, JSON, and git diff --check

External-data posture

No LBNL source archive was present locally. This PR records only original-fixture behavior and makes no full-data FPR, TPR, threshold-tuning, or card-validation claim. Users obtain data manually from the official source and follow its current license and redistribution terms. Full PFPU/SFPU mapping remains contingent on a reviewed local inventory-backed manifest and independently available evidence channels.