Three intentional patient-output deltas vs. migration:

1. Date typo rescue: clean/date_parser.py adds rescue_date_typos() and
   TYPO_REPLACEMENTS so common Excel-date typos (e.g., misspelled month
   strings) round-trip to a valid date instead of becoming Undefined.
   clean/converters.py wires the rescue through parse_date_column.
   errors.py registers the "typo_rescued" log code.

2. Multi-insulin CSV acceptance: clean/validators.py adds the
   allow_csv_subset flag so 2024+ trackers' comma-separated
   insulin_subtype values (e.g., "pre-mixed,rapid-acting") survive
   validation in canonical case instead of being replaced by Undefined.
   reference_data/validation_rules.yaml flips insulin_subtype on.
   clean/patient.py: docstring update on _derive_insulin_fields.

3. load_numeric_ranges helper + numeric_ranges: YAML block (consumed by
   the validate/ tree shipped in a later commit). Carrying both now keeps
   the YAML self-consistent.

Plus a small per-row dedup optimization in clean/converters.py and
the matching test updates.

Lays the groundwork the product cleaning pipeline (next commit) needs:

- extract/product.py: read the Stock_Summary section across month sheets
  (parallels extract/patient.py and reuses extract/common helpers).
- extract/wide_format.py: Mandalay-specific wide-format handling
  (column expansion 2020-21; cell splitting 2017-19).
- clean/schema_product.py: 19-column product meta schema + helpers.
- reference/products.py: known-product reference loader (categories,
  canonical names) backing product validation.

Reusable patient-side polish that the product layer now depends on:

- reference/synonyms.py: cache load_*_mapper with @lru_cache and fix the
  scalar-string -> list normalization so single-string YAML entries no
  longer iterate as characters.
- reference/loaders.py: explicit UTF-8 in load_yaml.
- tables/logs.py: schema_overrides on parse_log_file.
- clean/schema.py: docstring touch-up.

- clean/product.py implements R steps 2.0-2.21 from
  helper_product_data.R: synonym mapping, type conversion,
  patient-id repair, supplier-name preservation, units/balance
  reconciliation, end-row zeroing, and the per-sheet partition
  required for the .over() rolling balance to match R's per-sheet
  loop. Reuses safe_convert_column, fix_patient_id, and
  load_product_mapper from migration.
- tables/product.py aggregates cleaned product parquets into the
  single product_data table.
- tests/test_clean/test_product.py covers the cleaning steps;
  tests/test_tables/test_link_product_patient.py exercises the
  cross-product/patient link assumptions used by downstream tables.

Adds the manifest-driven incremental skip path that future commits will
expose behind --incremental:

- state/source.py + manifest.py: load the previous run's manifest from
  BigQuery, falling back to local parquet, then to an empty manifest.
- state/filter.py: skip trackers whose MD5 + per-pipeline completion
  flags match the previous manifest.
- tables/metadata.py: build tracker_metadata (MD5 + per-tracker output-
  presence flags) at table-creation time; md5_file is reused by filter.py.
- gcp/bigquery.py: add select_tracker_metadata and register product_data
  + tracker_metadata in PARQUET_TO_TABLE.
- state/__init__.py: additive re-exports for the new helpers.

Tests cover each state module, the metadata table builder, and the new
BigQuery accessor.

Wires the product cleaning + state machinery into a runnable pipeline:

- pipeline/product.py: orchestrates product extract+clean+table for the
  full corpus, mirroring pipeline/patient.py.
- pipeline/tracker.py: adds process_tracker_product so the per-tracker
  worker can run patient and product arms together.
- cli.py: introduces process-product, create-product-tables, and the
  unified run-pipeline command (patient + product + drive/GCS/BigQuery);
  adds --force and --incremental flags across process-patient,
  process-product, and run-pipeline. Imports the new state helpers and
  product entry points. The full ~810-line diff lands in one commit
  because the new commands and flags interleave with existing patient
  command bodies.
- tests/test_cli/test_force_incremental.py covers --force/--incremental
  for both process-* commands and run-pipeline.

Breaking change:

- tables/__init__.py drops the wildcard re-exports from .patient,
  .clinic, .logs, .metadata, .product. Callers must now import from the
  specific submodule (a4d.tables.patient, a4d.tables.product, ...).
  Verified zero affected call sites in src/, tests/, and scripts/.

Plus the doc/config drag-along:

- justfile: new product + run-pipeline recipes.
- readme.md, docs/CLAUDE.md, docs/migration/MIGRATION_GUIDE.md,
  docs/migration/PYTHON_IMPROVEMENTS.md: product pipeline + state docs.
- docs/PRODUCT_DATA_PIPELINE_FEATURE.md moved into docs/migration/.
- .gitignore: ignore .Rproj.user, .Rhistory, CLAUDE.local.md, and
  r-archive/config.yml.

Adds the source-vs-output validators that re-read each tracker's source
cells and compare against the cleaned parquet to flag silent data loss
or schema drift before tables are uploaded.

- validate/source_vs_output_patient.py + source_vs_output_product.py
  walk every patient/product row in the cleaned output, locate the
  matching source cell, and emit a structured diff (typed as
  match/converted/missing/unparseable). Activates the load_numeric_ranges
  helper and numeric_ranges: YAML block carried in the typo-rescue commit.
- validate/common.py holds the shared diff and IO primitives.
- Tests cover both patient and product paths against fixture trackers.

Read query rows directly and build the polars DataFrame with an
explicit schema instead of going through pandas. Removes a hidden
dependency on pandas (not declared in pyproject) and makes the
schema deterministic when the result set is empty.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Remove explicit encoding="utf-8" from open() calls and the
env_file_encoding setting (utf-8 is the default on the platforms
we target), and drop the Windows backslash-to-forward-slash
conversion in GCS blob names since Path.relative_to already
yields POSIX separators on macOS/Linux.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Removes the windows-shell directive, [unix]/[windows] recipe
attributes, and the [windows] PowerShell clean recipe. The project
targets macOS (dev) + Linux (Cloud Run / Docker prod); these shims
were drag-along noise from 85891e7 and were never present on
migration. Recipes that rely on bash heredocs will now fail loudly
on Windows, which is correct for an unsupported platform.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>