Dbt Agent Readiness

Precision follow-up to the 1.6.0 conditional-severity rework. 1.6.0 routes
dbt-pinned multi-home contradictions to adjudication (repo-grounded Blocker vs
metadata-grounded Hygiene); validation on the 13 public blog repos showed that
population was 618 candidates but almost entirely noise. The old home heuristic
counted a bare identifier-name occurrence as a definitional home — a column in a
runbook SQL query, a README file-format table, an onboarding checklist — and
short names produced false homonyms (cagov matched account_name / schema_name
/ credits_used against Snowflake admin SQL; gitlab stage against a "Snowflake
stage object" runbook note; email / username against a code-terminology
README list). Adjudicating 257 of the 618 found zero genuine docs-vs-dbt
contradictions.

• A doc HOMES an identifier only in a definitional context (docs_scan.py).
  _is_doc_home (a bare-token match against backticks, any table cell, or any
  fenced-code token) is replaced by _definitional_homes, which counts a home
  only when the doc DEFINES the term: the identifier is a heading's subject
  (## fct_orders), a column-dictionary row key, the subject of a "x
  is/means/represents …" prose definition, or a glossary entry. The glossary case
  is counted only in a glossary-typed doc — a colon-list bullet like - email: …
  in a README is code terminology, not a data home. Fenced code is stripped
  first, so a query selecting a column or a setup snippet exporting a variable is
  a use, never a home. A bare mention in prose, a backtick, or a
  checklist/metadata table cell no longer homes the identifier.
• Population shrunk from noise to signal. Across the 13 blog repos, total
  multi_home candidates dropped 633 → 116 and the dbt-pinned population that
  1.6.0 routes to adjudication dropped 618 → 116 (−82%); the no-fallback
  (Blocker-eligible, LLM-queue) population dropped 15 → 0. Per-repo (before →
  after): gitlab 41→4, cal-itp 82→4, tuva 250→37, stripe 90→0, full-funnel
  61→43, ga4 25→17, salesforce 21→0, snowplow 19→10, gtm-analytics 17→0, cagov
  11→0, jaffle_corp 8→0, jaffle_shop 7→0, mattermost 1→1. Every survivor
  inspected is a genuine definitional home — a column-dictionary table, a
  model-overview README table, or a metrics table — not a bare reference.
• docs_scan bumped to 1.5. The output shape is unchanged; only which mentions
  carry home: true changed, so multi_home_candidates, the LLM queue, and the
  dropped_beyond_cap counts all narrow to the high-signal set. Identifier
  coverage, column drift, external pointers, and staleness are computed from
  mentions (not homes) and are unaffected.

Pairs with 1.6.0: that change routes dbt-pinned multi-home to adjudication; this
change makes that routed population small and high-signal. The two are intended
to ship to the public repo together as one consolidated release.

Test suite 114 → 126 (12 home-precision assertions in tests/test_docs_scan.py
asserting that a fenced-SQL mention, a checklist-table cell, a code-terminology
colon-list, and a bare backtick are NOT homes while a heading subject,
column-dictionary row, prose definition, and glossary entry are; plus planted
test-fixtures/docs-context/home-precision/ fixtures). The public repo is
untouched.