grans label import: validate and ingest labeling corrections JSON

[dmwyatt]

Part of #66. Depends on storage (#69); pairs with export (#70).

Scope

grans label import <file.json> ingests the JSON downloaded from the labeling page into the speaker-label tables as label_source = 'human'.

Expected file shape (produced by the labeling template):

{
  "document_id": "...",
  "segmentation_version": 1,
  "labels": [
    {
      "utterance_id": "...",
      "speakers": ["person@example.com"],
      "segments": [{ "sentence": 0, "speakers": ["..."] }]
    }
  ]
}

segments is optional and present only for sentence-split labels.

Validation (strict; reject the whole file on failure, no silent skips)

• document_id and every utterance_id exist and belong together.
• Every speaker is a known candidate for that meeting, or unknown / media.
• segmentation_version matches the current one; on mismatch, refuse with a message explaining sentence indices may not line up (re-export and re-label rather than guess).
• Sentence indices within bounds for the utterance's current segmentation.

Behavior

• Upsert per utterance: importing again overwrites that utterance's previous human label (labeling sessions are iterative).
• Report a summary: N utterances labeled, M with segments, K speakers seen.
• Update README (new command).

[dmwyatt]

Scope narrowed by native attribution (see #66). Moving to status: triage.

What this now buys us. The least independently valuable issue in the set: it is the return path for #70, so it has no value unless #70 ships, and #70 is backfill-bound. Build it only as part of that chain.

One piece does generalize. The strict-validation stance (reject the whole file on failure, refuse on segmentation_version mismatch rather than guessing at sentence indices) is the right contract for any label ingestion, including corrections to Granola's native attribution if #69 grows a granola label source. If the correction path is built without the full backfill chain, this validation logic is the part worth lifting.