schema v0.1: add 'feature' to training_use

[rmems]

Goal

Add feature to the training_use enum in schemas/pr_trajectory.schema.json.

Why

Schema v0 has no feature value, so feature PRs are labeled other — flagged in
STATUS.md ("Remaining gaps") since the myelin extract. Six committed records carry
training_use: "other" with a feature card bucket today:

• corinth-canal #91, #95
• grok-ozempic Extract grok-ozempic-v0 trajectories (shortlist of 6 PRs) #11, data: refresh grok-ozempic shortlist post-v0 (#69–#79 + #42) #25
• myelin-accelerator normalize: dedupe review_signals and deprioritize 'Addressed in <sha>' acks #18, data: worktrees-hives agentic-workflow trajectories #26

Scope

• schemas/pr_trajectory.schema.json — add feature to the enum (backward-compatible;
  existing other records still validate)
• scripts/lib/normalize.py — FEATURE_BUCKET_TO_TRAINING["feature"] maps to feature;
  drop the now-redundant corinth TRAINING_USE_OVERRIDE entries that forced other
• tests/ — flip test_feature_bucket_maps_to_other_training_use to assert the new mapping
• Regenerate all 4 JSONL + manifests. This also absorbs the pre-existing drift found
  during refactor: make the repo parallel-safe for concurrent data extracts #37: corinth-canal-v0 and limen-axon-encoder-v0 no longer reproduce from their
  raw records (normalizer improved after their last regeneration). Drift absorption and
  the schema flip land as separate commits for reviewability.
• Add scripts/build_manifest.py so manifests are generated, not hand-authored —
  needed here to regenerate them verifiably, and by every v0.6 extract after.
• Update the schema note in docs/source-repos/_index.md and the card notes strings
  that mention the feature→other mapping; refresh the STATUS.md generated block.

Acceptance criteria

• The only JSONL changes are (a) the documented drift absorption in corinth/limen and
  (b) the six other → feature flips.
• build_manifest.py reproduces every manifest's structure; sha256/bytes match the
  regenerated JSONL.
• All tests green; validate_jsonl.py --strict-policy and build_status.py --check clean.

Blocks the v0.6 extract fan-out (#24/#25/#26/#27) — labeling must be consistent before
parallel extracts start.

[linear-code]

RM-633

[coderabbitai]

🔗 Related PRs

#12 - feat: read-only PR collector + corinth-canal-v0 trajectory JSONL [merged]
#17 - feat: grok-ozempic-v0 trajectories (revised 7-PR shortlist) [merged]
#35 - data: add grok-ozempic#42 Python trajectory to v0 (#20) [merged]
#36 - data: extract myelin-accelerator-v0 trajectories (#21) [merged]
#38 - refactor: de-conflict shared files for parallel extracts (#37) [open]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!