Add read_plate_dataset(): Parquet reader for the plate dataset

[vcjdeboer]

Before this PR leaves draft

• Re-run R CMD check --no-manual on the head commit. It was skipped for 54868e8, 400f7bc, 2915b2a and c261d83 on Vincent's word ("skip the tarball rebuild"); last pass was on e9906ed.
• Re-run the webR load check (library(seahtrue) and read_plate_dataset() under webR 0.6.0). It was skipped for 54868e8, 400f7bc, 2915b2a and c261d83; last pass was on e9906ed.
• The parser's plate_dataset_schema_v1.json is on swamp_seahtrue's main branch, and its sha256 still equals PLATE_DATASET_SCHEMA_V1_SHA256.

Summary

Adds an exported Parquet reader for the plate dataset, read_plate_dataset(). A plate dataset is the Parquet file the lab's .asyr parser writes for one plate, one row per well per tick.

The reader:

• Reads the file as data only. Nothing from the file is evaluated or deserialised. Arrow and R metadata in the file are ignored (use_arrow_metadata = FALSE), and every column is reduced to a plain atomic vector.

• Rejects an unknown or missing plate_dataset_schema_version. Only "1" is known.

• Refuses anything outside schema version 1:

  • missing, extra, reordered or mistyped columns;
  • values that should be constant per plate, well, measurement or tick but are not;
  • a non-finite value in any double column, or a non-finite O2 (mmHg) or pH after conversion;
  • anything other than 96 wells named A01..H12;
  • a missing well or tick;
  • a plate id that is not a lowercase sha256;
  • a date_run that is not ISO 8601 UTC.

  File size, row count, uncompressed size and string length are all capped. Refusal messages quote file values only truncated and escaped.

• Converts emissions. O2 emission becomes O2 (mmHg) by Stern-Volmer with F0 and the uncorrected Ksv. pH emission becomes pH on a line through the calibration pH at the pH target emission.

• Keeps every well. flagged_well and plate_flagged_well are manual flags a person set while looking at the data, both carried as data. Their one effect: a Background well flagged in either is left out of the background average. When every Background well is flagged, or there is none, the background columns are missing.

• Returns seahtrue's plate structure with our names and runs validate_preprocessed() itself.

  • assay_info carries plate_dataset_schema_version and plate_dataset_schema_sha256, so callers can check which schema the plate was read with without using internals.
  • rate_data has zero rows: no OCR or ECAR yet.

• Adds nanoparquet (>= 0.5.1) to Imports. It has no dependencies and runs under webR; arrow does not.

The schema is a copy of the parser's plate_dataset_schema_v1.json: 34 columns, sha256 8ffe7fdd…4c5b, parser commit f6d95e1. The parser is still in review, so this PR stays a draft.

Checks

• devtools::test() at c261d83: 171 expectations pass, 0 failures, 89 of them for the new reader on synthetic plate datasets. No real plate data is in this repository.
• At e9906ed:
  • R CMD check --no-manual: 0 errors, 0 warnings, 2 notes. One note is new: Imports now has 21 non-default packages.
  • webR 0.6.0 (R 4.6.0): the package loads with library(seahtrue) and the reader works.
  • Both real plates the parser produced then read and passed validation locally.
• Both checks are still to be re-run for c261d83 (see the checklist at the top).

Do not merge or tag from this PR; Vincent decides the release.

🤖 Generated with Claude Code

[vcjdeboer]

Revise round (Vincent at acceptance: "4 revise, F2 ok"), commit 54868e8:

• Non-finite values are refused: Inf, -Inf or NaN in any double column, and a non-finite O2 (mmHg) or pH after conversion. A test covers both.
• Tests now cover the footer row-count cap, rows read differing from the footer, and the uncompressed-size cap, using a mocked footer reader.
• The schema v1 copy follows the parser's updated file (sha256 8ffe7fdd…, parser commit f6d95e1). It adds a required BOOLEAN column, plate_flagged_well. Both manual well flags are carried as data: no well is dropped or excluded, and background wells are averaged whether flagged or not.

devtools::test(): 157 expectations, 0 failures.

On Vincent's word ("skip the tarball rebuild"), R CMD check and the webR load check were not re-run this round. Both must be re-run once before this PR leaves draft.

🤖 Generated with Claude Code