Independent analysis batches

The batch coordinator keeps ordered independent inputs and statuses outside the Assay and Sequence scientific schemas. analysis-batch@1 is a local project kind; it is not a pooled assay or a new numerical method. Every child uses the existing selected analysis, reporting refusals and method versions. Comparison never treats plates or fitted observations as independent biological experiments.

The common envelope reader is in src/lib/batch/model.ts. App adapters must also validate scientific child payloads. Unknown versions and fields are refused; refused, cancelled and not-run items remain part of the recoverable batch.

Implementation and final verification are recorded in docs/progress/batch-2026-10-16.md during integration.

Inputs and independent analysis

The lazy Assay panel receives a frozen reviewed import and a scope/epoch captured at the explicit transfer action. Editing the original import box does not alter that copy. The researcher confirms that blocks are separate dose-response plates and chooses a compatible reviewed map per plate. Reporter pairs and kinetic reads are refused. Each child gets a fresh source ID, one plate, remapped review locations and exclusions, and no previous results or experiment declarations. The existing deriveCurves, sequential FitClient, summarise, method pins and dose-response input hash produce each output. Display uses shownLogIC50 and shownInterval; counts are fitted observations. No across-plate statistics are computed.

The Sequence adapter compares parsed record counts against FASTA/GenBank headers, retains complete source text and exact per-record text, and uses the shared loadedRecord conversion. The existing selected operation runs independently on each record. Worker-sized six-frame, ORF and motif analyses use the existing worker client sequentially. sheetRows supplies the values, units and method versions. Empty, ambiguous, incompatible and operation-refused inputs remain visible. Primer design is explicitly unavailable in this batch path.

The bounds are 24 plates, 4 MiB of original import/text and 400 derived curves for Assay; 100 records and 1,000,000 UTF-8 source bytes for Sequence. A shared Sequence template keeps its existing separate 1,000,000-byte bound. Bounds reject the whole batch before computation, without truncation. Comparison cells preview the first 1,000 characters of long outputs; recovery retains the full output.

Lifetime, cancellation and recovery

The queue captures exact input/settings identity, scope ID, scope epoch and a run generation. It yields between children, checks after every asynchronous boundary, and disposes workers on input replacement, scope changes or unmount. UI action lifetimes also guard file reads, services and transactional writes. Previous-scope inputs, comparison cells and saved-batch lists are hidden. An input/settings edit retains completed cells as stale and marks an interrupted row cancelled. It cannot record those outputs as current.

User cancellation preserves completed children, cancels the interrupted child and leaves subsequent children not run. Resume retries failed, cancelled and not-run inputs; it retains completed and refused rows. A fresh run recomputes all children. Opening a saved or imported batch shows previous outputs as stale and requires a fresh run before recording. Historical values are not silently recomputed during recovery admission.

createBatchStore explicitly validates before both project creation and revision-checked save; createProjectStore.create alone is insufficient. Its fixed-scope transaction wrapper also checks the actual epoch during writes and prevents an unsafe independent quota fallback. The common reader, exact input hash and app readers reject malformed versions, altered input membership and inconsistent output/payload/row relationships. Assay retained outputs also pass the existing dose-response input-integrity check. Sequence outputs are retained as historical values until rerun; this is not a new independent numerical verification engine.

Recovery uses the existing project-export envelope with kind analysis-batch and schema 1. It includes every input/status, sequence text or plate reading, source context and per-child output. It excludes original instrument-file bytes; file name/hash/size remain provenance with retention explicitly false. Restoring creates a new local project and removes local child/result receipts and planned write identities. Unknown schema versions are refused before a write. Memory fallback is stated explicitly and users can export recovery.

Recording reuses saveObjectBackedWork, the existing app capture adapter and snapshot store with default stale rejection. Each child commits independently. The coordinator checkpoints planned object IDs and prepared input/result IDs before their corresponding write, then records receipts. A retry reuses committed identities, including an interruption after a result committed but before its receipt checkpoint. Completed children are unfiled Bench work; batch commands do not implicitly assign Study membership. This is partial, per-child saving, not an atomic batch transaction. All services work in the browser-local workspace with no account or remote backup dependency.