Authoring-assist: suppressor signals, prose mining, applicability + keyword checks

Expand Suggest-ArticleSignals.ps1 from signals-only to the full routing-relevant
head-matter:
- Suppressor detection: 'not-a-violation' articles (by title/prose, excluding
  prohibitions) propose effect: suppress mapping-form signals.
- Prose mining: extract AL constructs from inline code + fenced al blocks so the
  ~17% of sample-less articles are covered (marked proseOnly); -NoProse opts out.
- Applicability cross-check: flag JS/control-add-in use when technologies omits
  javascript.
- Keyword suggestions: strong proposed tokens missing from keywords.
Generated suppress blocks pass validate_frontmatter R24. Suggest-only, offline.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

tools/authoring-assist.md

@@ -20,17 +20,37 @@ critical work we don't want to guess at. This tool lowers that authoring cost.
 ## What it does
 
 For each article it reads the article's own co-located `good`/`bad` `.al` samples
-plus its front-matter, extracts candidate AL constructs, and emits:
+**and its prose** (inline `` `code` `` spans + fenced ```` ```al ```` blocks), plus
+its front-matter, extracts candidate AL constructs, and emits:
 
 1. **Proposed `signals:` block** — the highest-value constructs that are *not*
    already routed to the article's domain by `routing-seed.json`. Constructs are
    ranked by: keyword-backed (+3), appears in the `.bad.al` anti-pattern sample
-   (+2), frequency. Sample-local scaffolding (procedure/object names) and
-   structural AL vocabulary (types, control-flow, object kinds) are filtered out.
-2. **Domain-mismatch flag** — when ≥2 of an article's sample constructs are known
+   (+2), seen in an actual sample (+1), frequency. Sample-local scaffolding
+   (procedure/object names) and structural AL vocabulary (types, control-flow,
+   object kinds) are filtered out.
+2. **Suppressor (`effect: suppress`) proposals** — when an article exists to say a
+   construct is **not** a violation of its domain (detected from the title/basename
+   and prose, e.g. `page-display-is-not-a-privacy-concern`), its constructs are
+   proposed as `effect: suppress` (mapping form) instead of raise. A suppress hit
+   *dampens* the domain score in the orchestrator rather than raising it, directly
+   attacking the over-firing (precision) problem. Prohibitions (`do-not-…`,
+   `avoid-…`) are explicitly **excluded** — those describe a real violation.
+3. **Domain-mismatch flag** — when ≥2 of an article's sample constructs are known
    to the seed and they route to a *different* domain than the article declares
    (e.g. an `events/` article whose sample is all `TryFunction`/`Error`). A soft
    advisory prompting the author to confirm the domain or add a cross-domain signal.
+   Skipped for suppressor articles, which deliberately reference other domains.
+4. **Applicability note** — flags samples/prose that use a JavaScript control add-in
+   when `technologies` omits `javascript`. Small but concrete; kept minimal because
+   the technologies vocabulary is tiny (`al`/`javascript`).
+5. **Keyword suggestions** — strong proposed constructs not already present in the
+   article's `keywords` list, surfaced as an advisory "consider adding".
+
+**Prose mining** matters for the ~17% of articles that ship no `.al` samples at all
+(including most suppressor articles): their constructs are recovered from the prose
+and marked `proseOnly` so a reviewer knows the weaker provenance. Use `-NoProse` to
+restrict the source to samples only.
 
 It is **deterministic, offline, and open**: no feedback/telemetry data, so the
 community can run it against `community/` and `custom/` layers too.
@@ -45,15 +65,18 @@ tools/Suggest-ArticleSignals.ps1
 tools/Suggest-ArticleSignals.ps1 -Path knowledge/style
 tools/Suggest-ArticleSignals.ps1 -Path avoid-commit-inside-loops
 
-# Only the high-value cases (zero-signal domains + mismatch flags)
+# Only the high-value cases (zero-signal domains, suppressors, mismatch/applicability)
 tools/Suggest-ArticleSignals.ps1 -OnlyGaps
 
+# Samples-only (ignore article prose)
+tools/Suggest-ArticleSignals.ps1 -NoProse
+
 # Machine-readable (for a future automated PR advisory to consume)
 tools/Suggest-ArticleSignals.ps1 -AsJson
 ```
 
-Key parameters: `-Top <n>` (max signals per article, default 3), `-BCQualityRoot`,
-`-SeedPath`.
+Key parameters: `-Top <n>` (max signals per article, default 3), `-NoProse`,
+`-BCQualityRoot`, `-SeedPath`.
 
 ## Guardrails (why it only suggests)
 
@@ -61,7 +84,8 @@ Key parameters: `-Top <n>` (max signals per article, default 3), `-BCQualityRoot
 would inject the same noise we're trying to remove and would bypass CODEOWNERS
 review. So the tool emits a proposal an author eyeballs and adds via a normal PR;
 [`validate_frontmatter.py` R24](../.github/scripts/validate_frontmatter.py) + CI +
-human review remain the gate. Proposed bare-token blocks satisfy R24 as-is.
+human review remain the gate. Both proposed bare-token (raise) and mapping-form
+(`effect: suppress`) blocks satisfy R24 as-is.
 
 ## Relationship to the feedback-driven engine (later, Tier-2)