feat: add native prospective instrumentation

[jimmlucas]

What changed

• adds a native esdp init command for prospective long-read instrumentation projects
• adds esdp record-round to collect one light observation and rebuild its causal history as one immutable round transaction
• exposes the same workflow through Python init_project and record_round functions
• records sequencing technology, chemistry, ONT basecaller/model, assembler, polisher, versions, and frozen parameters
• generates a strict offline-QC JSON Schema for BUSCO counts/provenance, independent k-mer QV, separated consensus errors, homopolymer indels, and coding-integrity outcomes
• binds observations, external QC, reads, and project contracts through SHA-256 identities
• rejects skipped/overwritten rounds, unsafe identifiers, mismatched assembly QC, pre-polish alignment references, duplicate reads, and project-contract drift
• keeps every project and round permanently at decision_enabled: false
• packages the instrumentation API in the wheel and Docker runtime

Why

ESDP needs prospectively collected, provenance-complete long-read trajectories before another stopping model can be scientifically selected. This layer lets ESDP collect those data directly, without nf-core/Nextflow and without enabling experimental STOP/CONTINUE decisions.

Expensive evaluators remain external for now. ESDP validates and archives their results rather than silently introducing BUSCO, Merqury, alignment, or gene-prediction runtime dependencies.

Immutable project layout

Each trajectory stores independent R1...Rn directories. A round includes its observation, cumulative causal history, transaction record, and optional offline QC. Round directories are committed by atomic rename; recording R2 never modifies R1.

Every round records the SHA-256 of esdp-project.json. Changing technology or tool provenance after R1 blocks subsequent rounds.

User interfaces

esdp init --project-directory study ...
esdp record-round --project-directory study --sample-id isolate-1 --round 1 ...

Equivalent Python functions are documented in docs/PROSPECTIVE_INSTRUMENTATION.md.

Scientific boundaries

• no model is loaded or trained
• no stopping decision is produced
• expensive QC outcomes never enter the light feature history
• the held-out scientific test set remains untouched
• future collectors can execute external tools only after their commands, containers, cost, and portability contracts are frozen

Validation

• full suite: 201 passed
• instrumentation/CLI/integration focused suite: 63 passed
• wheel built successfully without build isolation
• wheel contains esdp_instrumentation.py, CLI, metrics, and history modules
• wheel installed into an external virtual environment
• installed esdp init successfully initialized a PacBio HiFi project outside the repository
• git diff --cached --check: clean

This is a stacked draft PR targeting feat/esdp-light so its diff contains only prospective instrumentation.