feat(courseware): extract each distinct scope text once and reuse it

[blarghmatey]

Stacked on #3038, for the task "Content-address courseware extraction so B2B course copies reuse one extraction". extraction_key hashes the platform, run and scope id, and B2B copies of a course share its text under new block ids, so every copy was sent to the model again. The local lake has 107,205 scopes holding 43,885 distinct scope texts (1.08 of 2.74 billion source characters), so about 60% of extraction input is copies.

This adds content_key, which hashes what a call reads apart from the course header (extractor id, prompt version, schema version, model, effort, scope kind and title, part and input_text_checksum) and leaves out the run and scope id. extract_scopes sends one call per content key, or takes a stored extraction with that key from the new reusable argument. Each copy still gets its own extraction row under its own extraction_key (stop_reason = 'reused', zero tokens) and its own claim and evidence rows, so covered_keys, rekey and the dbt models work unchanged (only two column descriptions changed). Evidence takes the copy's own source_ref by source id, since the model sees S01, S02 and not block ids, and verification and offsets carry over because the text is identical. A failed call leaves its copies out, so the next run retries them. local extract passes its stored tables as reusable, and the dry run marks copies reused.

A copy records the original's header_version, not its own, because its claims were made under the original's header and covered_keys hashes header_version into the header-keyed key. With its own header, a later --key-on-header run would count the copy as already extracted under that header and skip it. With the original's, is_header_stale from #3038 flags the copy and --key-on-header re-extracts it.

Measured with dry runs: xPRO MLx2 R14 to R16 reuses 190 of 288 scope parts and sends ~450k of ~1.28M input tokens. The task's figure of 6,224 runs holding 3,260 distinct texts didn't reproduce; the local lake has 6,222 distinct run checksums with ids and 5,823 without.

Testing: 279 ml tests pass, including local extract run twice over a run and its copy (one model call, the copy stored with the original's header and its own claims, evidence and scope_input row) and a --key-on-header run re-extracting a copy made under another header. ol-dbt unit-test passes.

Open questions for review:

• Reuse ignores the course header across courses, consistent with the 2026-10-17 decision to keep the header out of the key. That means identical boilerplate in unrelated courses shares claims.
• An untitled source's block id is its title in the message, which keeps about 7,000 more scopes from deduplicating. Fixing that changes the message, so it needs an EXTRACTOR_ID bump.

Not covered: the Dagster extraction asset in #3035 doesn't pass stored rows as reusable, so production gets none of the savings until it does.

https://claude.ai/code/session_01QJTJTyry3Ep28ndqEfqUgK

[github-actions]

🔎 ol-dbt impact — column-level blast radius

✅ No column-level downstream impact detected for the changed models.

_{Posted by ol-dbt impact (annotate-only — does not block merge).}