What this is

MC-30 is now scoped to the documentation side only. This PR does two things:

1. Updates docs/design/Database-entity-details.md to match five changes that
   have landed on the pipeline side, so the entity and field tables describe
   what the pipeline actually emits.
2. Deletes docs/design/schemas/pipeline-records.schema.json. The schema
   arrives with the POC-3 migration ([MC-1] NACHOS automated scoring engine - initial code migration #11), which brings in the generator that
   produces it.

One commit each.

1. The doc sync

Fact-run identity — fact_runs gains the dedup key

A fact-run header carries mode, model, prompt_version, status,
cost_cap_hit and the run counts. Nothing the landing side deduped on covered
any of it: ingestion_runs.run_id comes purely from the ingestion inputs, and
the per-(state, lens) release_id covers scored records.

That gap was observed, not theorised. Correcting fact_run.mode from "api" to
"llm" (MC-24 §8.8) changed 135 fact-run headers while every identity the loader
compares stayed byte-identical. The load concluded the run was already landed and
skipped the correction.

Two digest columns close it (nachos-ai-poc-3#345, ADR 0023):

• observations_digest — sha256 over the observation rows. What was extracted.
• fact_run_digest — sha256 over the identity-recipe version, every
  non-volatile header field, and observations_digest. The dedup key. Equal
  means the held row is current, skip. Different means re-land, including when
  ingestion_run_id and the snapshot digests are unchanged — the case that
  broke. Comparing the pair separates a header-only correction (observations
  equal, run digest moved) from changed data (both moved).

The digest does not cover scored_at or the token/cost/cache counters: those
describe the occurrence, and a cache-warm replay of identical facts reports $0
and different token counts without being new data.

Adds a Fact-run identity section plus the two fact_runs rows.

base_nachos_score on score_records

adjusted_nachos_score was already a column; the base methodology tier it is
built from (0–3) was reachable only inside the open dimensions blob
(dimensions.nachos_score.value), so one half of the base/adjusted headline was
a plain column and the other needed a JSON dig.

Landing-derived from dimensions.nachos_score.value the same way
documentation_gap is, so nothing changes on the emit side.

Distinct from complexity_score: nachos_score is the NACHOS methodology tier
that feeds adjusted_nachos_score; complexity_score is the
business_logic_complexity comprehension-cost signal. Same input facts,
different rubric — they coincide on ~99% of rows but diverge where the
methodology rules differ: 238 / 24,990 rows in release r-37334d4a830f060d.

source_document_description and school_year on source_elements

Operator-supplied provenance for the documentation package an artifact was built
from — a plain-English label like "TEA publication Jan 2026" and, optionally, the
school year it targets — sourced from the pipeline's committed
data/source_documents.yaml registry. Both are uniform across every record of one
artifact; display-only, never scored. source_document_description is required
non-blank, school_year is required-but-nullable.

nachos-ai-poc-3#351 (ADR 0016).

documented on score_records; sidecar contract 3 → 4

documented is the headline predicate. Aggregate computes every reported mean,
dimension distribution and review count over the rows where it is true (the
header's in_scope_count is their count), but no per-record field let a consumer
reproduce that population from a sidecar alone. documentation_source looks like
it should serve and does not: on the API-model lens every row is source_doc, so
a consumer keying on it computes the headline population correctly on one lens and
silently wrongly on the other.

Per ADR 0017 the contract-version bump does not stale adjudications.

nachos-ai-poc-3#364 (ADR 0025).

prompt_fingerprint and answered_by_host on fact_runs; fact identity 1 → 2

• prompt_fingerprint — sha256 over the rendered prompt batches that produced the
  observations; the producer's own freshness key, which freshness previously could
  not detect a change to.
• answered_by_host — which endpoint served the run (anthropic,
  azure:<host>/<deployment>, deterministic, the cache:/mixed: replay tags,
  or unknown). Records what model cannot: an Azure-hosted run still reports
  the base model id.

Both are nullable and covered by the ADR 0023 deny-list by construction, so the
digest recipe is unchanged; the fact-identity version moves 1 → 2 purely as an
epoch marker. One-time consequence: every fact_run_digest in the tree changes at
once while every observations_digest stays put, so a loader re-lands each fact
run once rather than reading independent content changes.

This also corrects the fact_runs.contract_version row, which described the
sidecar meaning — the exact touchdownllc/dataload_poc#27 conflation. A fact run has
no sidecar; its contract_version is the fact-identity recipe epoch.

nachos-ai-poc-3#378/#381 (ADR 0027); host tags #392, refined by #394.

2. Deleting the vendored schema

docs/design/schemas/pipeline-records.schema.json was a second copy of an artifact
the pipeline generates and owns (publish/pipeline_records_contract.py), kept
byte-identical by hand because the generator and the consumer of the contract lived
in separate repos. The POC-3 migration brings the generator into this repo, so the
copy stops being a sync target — it would be a stale sibling of a file generated a
directory away.

Database-entity-details.md pointed at it by repo path in five places. Those now
name the companion schema without pinning a path, since the migration settles where
it lands, and the two-repo byte-identity discipline goes with the copy.

MC-18 is unaffected in substance. The contract test still validates
pipeline-emitted records against the schema on every build; what it no longer has is
a second copy to enforce byte identity against. That is a simplification of MC-18,
not a gap — it should be picked up when the migration lands and the schema's path in
this repo is known.

Files

• docs/design/Database-entity-details.md — the five changes above, plus the
  path-free rewording of the five schema references.
• docs/design/schemas/pipeline-records.schema.json — deleted.