feat(field-observation): ingest the 42 + 38 attributed @nvst18 lots

[waitdeadai]

Ingest the two attributed @nvst18 field-observation lots (42 + 38)

Supersedes the scaffold #19. Both contributor gates cleared on 2026-05-31
(@nvst18 sign-off on
ianymu/recognition-without-arrest#2:
"The 42 (redaction): Approved… The 38: Generalize the internal class/method
names… before ingest. Both lots good to ingest once the placeholders are in."),
so this materializes the payloads the scaffold was holding for.

What lands

Two separately-tagged sub-batches in one attributed lane (distinct
source_id, sub-batch dir, and manifest transform — provenance never merges):

Sub-batch	Records	category	Handling
dispatch_fabrication_42	42 (39 pos / 3 neg)	sycophancy	Ingested verbatim from the signed-off redacted candidate. Agent codenames + "Want me to dispatch X" phrasing preserved as the fabrication signal, per your approval.
hollow_code_38	38 (35 pos / 3 neg)	hollow_code	Internal class/method/const/metric identifiers generalized to role-preserving surrogates (e.g. NaomiGuardrail→ContentGuardrail, getCrisisText→getSafetyText, ScorePtgiWeekly→ScoreWeekly, ptgi_weekly→score_weekly).

Plus a derived_fixture_manifest.jsonl (80 rows, one per record, hashing each
original source row), registry/README/LANE_SCHEMA updates, and lane registration
in DATASET_CARD.md + CLAIM_LEDGER.md.

Generalization approach (the 38)

Type-preserving pseudonymization: ~30 distinct identifiers mapped to neutral
surrogates that keep the code syntactically valid (so the fixtures stay
usable for detector training) and preserve the failure signal — e.g. the
safety_prompt_bypass ->body vs getSystemPrompt() contrast survives intact.
A hard residual gate asserts zero original proprietary token (Naomi,
PTGI, Crisis, Therapist, Externalization, NarrativeLetter,
Reflection, Disengagement, and every original class name) survives anywhere.

Two notes for your review @nvst18

1. No commit SHAs were stripped because there are none in the 38 — every
   fix_description is a short code-fix note (e.g. "Captured $validated and
   used $validated['rating']"), not a SHA. (The 27af4ec ref lives only in the
   registry prose as a fix example.)
2. The 42 (your approved redaction) is verbatim. A maintainer reject-scan
   surfaced only items inside your sign-off — short session-id hex (e.g.
   a2c8f421 = "Session a2c8f421") and a generic github.com/topics link. I did
   not alter your signed-off content; flagging in case you'd like the
   session-id prefixes stripped before this merges.

Status / safety

• label_final = null on every record (candidate only; not a released
  data/ lane until two annotation passes + adjudication).
• Validation run: registry/manifest/payloads parse; validate_corpus.py clean;
  release_workflow_safety_check.py --scan-secrets exit 0 / errors []; pytest
  20 passed; reject-to-quarantine scan clean on both sub-batches.

Attribution to @nvst18 is retained per the contribution terms. Requesting
your review as the contributor before this merges.

🤖 Generated with Claude Code