- Skip the template value-type descriptor row during ingestion (both etcher
  and generic builders) so a filled-in template processes cleanly.
- Attribute generic runs to their own equipment_id and resolve the display
  name via equipment_metadata instead of the project's equipment.
- Surface generic measured outputs in the catalog run-detail page and render
  generic input set-points (strings or numbers); add outputs to V2Run.
- Count generic equipment_runs in equipment inventory run metrics.

Co-authored-by: Cursor <cursoragent@cursor.com>

- Explicit-column templates now also include lot/timestamp keys and registered
  input parameters so no run metadata is silently dropped.
- process_upload rejects linking a run to a project whose equipment differs
  from the upload's equipment (empty project equipment stays unconstrained).
- run_to_jsonld serializes generic measured outputs and labels the dataset by
  its actual equipment instead of always "Etcher".

Co-authored-by: Cursor <cursoragent@cursor.com>

- _coerce_optional_float rejects NaN/Infinity so generic ingestion stores the
  raw token as text instead of emitting invalid JSONB and 500-ing.
- _fetch_equipment_run_metrics aggregates on the superuser connection so
  inventory run counts/last-data are complete fleet-wide rather than being
  hidden by RLS for private/shared projects.

Co-authored-by: Cursor <cursoragent@cursor.com>

- Scope the catalog Download button explicitly to the etcher ML export
  (label + tooltip + filename) since that endpoint is the model-specific
  training matrix; generic runs remain browsable in the catalog.
- Validate generic run timestamps and required lot/timestamp cells during
  processing so malformed dates surface as actionable validation errors
  instead of a database-sync 500; store normalized ISO timestamps.

Co-authored-by: Cursor <cursoragent@cursor.com>

- Required upload columns are now the equipment's declared inputs (explicit
  columns, else features+parameters); outputs/targets and synthetic
  placeholder targets (primary_metric/secondary_metric) are never required,
  so valid CSVs for newly registered equipment are no longer rejected.
- Generic ingestion coerces each cell by its registered type, preserving
  zero-padded string ids, integers, and booleans instead of floating all.
- run_to_jsonld / _load_domain_config fall back to the equipment's PG
  config_json so FAIR metadata for PG-registered equipment carries units,
  ontology, hardware, and owner info.

Co-authored-by: Cursor <cursoragent@cursor.com>

- equipment_runs gains a stable (upload_id, row_index) identity; reprocessing
  an upload now upserts in place (preserving each run's id/run_id and catalog
  URL) and prunes only rows dropped by a shorter re-upload.
- _load_domain_config returns the file config only when present so the PG
  config_json fallback actually runs for database-only equipment.
- _get_feature_ontology_map now indexes registered outputs so their units and
  QUDT metadata appear in FAIR JSON-LD.
- Run-detail page renders boolean/string generic measurements explicitly
  instead of rendering nothing.

Co-authored-by: Cursor <cursoragent@cursor.com>

- run_to_jsonld selects the etcher target fallback by equipment_id, so a
  non-etcher run with no outputs no longer emits null AvgEtchRate/RangeEtchRate
  properties and is correctly labeled.
- Template descriptor-row detection now requires the complete hint row and is
  only applied to the first data row, so a sparse data row that happens to equal
  a type token is never silently dropped.

Co-authored-by: Cursor <cursoragent@cursor.com>

_required_upload_columns now always requires the lot and timestamp metadata
columns plus declared inputs (features + parameters), and excludes outputs even
when a legacy config lists them as explicit columns. This stops anonymous/
undated catalog runs for new equipment and stops wrongly requiring output
columns (or omitting appended parameters) for explicit-column configs.

Co-authored-by: Cursor <cursoragent@cursor.com>

- upload_file validates registered equipment against _required_upload_columns
  always (even with explicit columns), so an inputs-only CSV is processable.
- Generic ingestion reports an error for non-empty cells that cannot match
  their declared numeric/boolean type instead of silently storing bad data.
- Template descriptor-row detection only compares hint columns present in the
  uploaded row, so a trimmed inputs-only template is still recognized.
- Range warnings recorded at validation time are carried through processing
  (errors_json) and returned, keeping the processed-upload warning UI working.

Co-authored-by: Cursor <cursoragent@cursor.com>

- upload_file validates canonical-etcher CSVs against the full alias-aware
  REQUIRED_SYNC_COLUMNS (including outputs) so an upload accepted at validation
  time can actually be processed by _csv_upload_to_sync_rows.
- process_upload always returns errors as an array (empty when warning-free),
  restoring the API contract the upload page relies on (result.errors.length).

Co-authored-by: Cursor <cursoragent@cursor.com>

_declared_type_violation now flags non-integral values (e.g. 1.5) for columns
declared int/integer/long, so fractional data can't be stored in fields the
equipment schema marks as integers.

Co-authored-by: Cursor <cursoragent@cursor.com>

- Template value-type hints now honor declared parameter/output types
  (string/boolean/int), so a non-float field is no longer advertised as float.
- etcher_runs always report equipment_id/name as the canonical etcher (even
  when the project has no equipment association), so FAIR JSON-LD keeps etch-rate
  results and etcher labeling instead of degrading to a generic run.

Co-authored-by: Cursor <cursoragent@cursor.com>

…utput fixes

- etcher_runs are unconditionally identified/aggregated as the canonical etcher
  (data_loader_pg base select and metadata_pg metrics), so a project associated
  with another tool (or none) can't misattribute or drop canonical runs.
- Template hints prefer a non-float declared type over a unit, so boolean/int/
  string fields advertise their real type while numeric fields keep unit hints.
- Generic ingestion omits None-valued input/output cells, so inputs-only or
  partially-filled uploads don't store misleading {"result": null} measurements.

Co-authored-by: Cursor <cursoragent@cursor.com>

get_summary_stats_pg now aggregates etcher_runs together with equipment_runs
for total/clean/outlier counts and the date range, so processing a non-etcher
upload updates the dashboard's run statistics instead of leaving them unchanged.

Co-authored-by: Cursor <cursoragent@cursor.com>

_get_feature_ontology_map now indexes registered input parameters (unit, QUDT,
prov_direction: input), so generic equipment that declare inputs under
parameters rather than features keep their units and provenance in FAIR JSON-LD.

Co-authored-by: Cursor <cursoragent@cursor.com>

…etadata

PG-registered equipment store manufacturer/model/serial_number/location at the
config top level, but run_to_jsonld reads them from config["domain"].
_load_domain_config now normalizes the PG config so these hardware/identity
fields are backfilled into domain, keeping them in generic-run FAIR JSON-LD.

Co-authored-by: Cursor <cursoragent@cursor.com>

…jection

- Merged etcher/equipment run pages sort on the actual instant (offset-aware
  datetimes normalized to UTC) instead of raw ISO strings, so runs with
  different UTC offsets paginate in true chronological order.
- Generic ingestion rejects structurally malformed rows (extra cells under the
  DictReader None key, or short rows leaving columns at the None restval) so a
  directly-invoked process call can't turn a malformed CSV into a catalog run.

Co-authored-by: Cursor <cursoragent@cursor.com>

…round-trip

Add _coerce_optional_int and use it for int/integer/long columns in both
_coerce_typed and _declared_type_violation, so 64-bit identifiers beyond
2**53 keep full precision instead of being silently rounded via float.

Co-authored-by: Cursor <cursoragent@cursor.com>

…nfigs

- _coerce_optional_int parses decimal-form integers (e.g. 9007199254740993.0)
  via Decimal instead of binary float, preserving 64-bit values; non-finite
  Decimals are rejected.
- _load_domain_config normalizes file-backed/registered JSON configs too, so
  manufacturer/model/serial/location stored at the config top level appear in
  FAIR metadata for equipment that have a JSON snapshot, not just PG-only ones.

Co-authored-by: Cursor <cursoragent@cursor.com>

…ions

Add _effective_equipment_config to backfill a registered equipment's explicit
column schema from its top-level columns (columns_json) when config_json is
empty/incomplete. upload_file and _load_equipment_config both use it, so the
upload schema validation and generic processing no longer accept or drop files
that omit registered columns. Restores test_upload_reports_missing_required_columns.

Co-authored-by: Cursor <cursoragent@cursor.com>

Merged run pagination now returns an empty page for limit=0 (matching the prior
SQL LIMIT 0 semantics) instead of treating a falsy limit as unlimited-from-
offset; limit=None remains unlimited and limit>0 returns the window.

Co-authored-by: Cursor <cursoragent@cursor.com>

Generic ingestion now parses is_outlier, is_calibration_recipe, and outlier_type
(via the standard column aliases) into the top-level row passed to
sync_equipment_runs_pg, so uploaded generic runs keep their quality metadata
instead of defaulting to false/empty and being misclassified in catalog filters,
badges, and summary outlier counts.

Co-authored-by: Cursor <cursoragent@cursor.com>