feat(agents): pre-PR build+test gate + tighter current-context lint

.agents/routines/context-refresh-prompt.md

@@ -70,6 +70,13 @@ routine's context.
 - Link every entry (`#1234`, `PR adcp-client#456`)
 - Status values: `active`, `blocked`, `review`, `shipped`, `deferred`
 - Drop stale entries — staleness is a feature
+- **Bullet labels must be factual nouns.** Not "**X gaps**",
+  "**X risks**", "**X concerns**", "**Tier-N X**", or
+  "**Stakeholder X**". Say what the work *is*, not how it's
+  framed. Example: "**Compliance storyboard remediation**", not
+  "**Compliance storyboard gaps**". The lint flags these, and
+  any gap/risk/concern/tier framing belongs in
+  `internal-context.md` instead.
 
 ## Open a PR
 

.agents/routines/triage-prompt.md

@@ -424,9 +424,35 @@ A single cohesive PR of 200 non-breaking lines is easier to review
 than three PRs of 60 lines with dependencies and cross-links. The
 bot's job is to reduce maintainer clicks, not multiply them.
 
+## Pre-PR build + test gate — mandatory before expert review
+
+The expert review is expensive; don't run it on broken code. Before
+spawning experts, make sure the diff actually compiles and the
+unit tests pass.
+
+1. Run the repo's build + fast test tier:
+   - `npm run precommit` — prefer this (bundled build + typecheck +
+     unit tests); falls back to `npm run build && npm test` if not
+     defined
+   - If the diff touches only `docs/**` or `static/**`, skip build
+     and run the relevant doc check instead (e.g., `npm run
+     docs:check` or `mintlify broken-links`)
+2. **If build or tests fail:** read the errors, fix the code,
+   re-run. Cap at **2 build→fix iterations.** If still failing,
+   abandon the PR and Flag for human review with the build log
+   in the comment.
+3. Do **not** skip tests locally because "CI will run them." The
+   point of this gate is to not ship known-broken code even as a
+   draft, because (a) review noise, (b) a human reviewer may
+   admin-merge a draft that looks fine, (c) a green CI on push
+   is the baseline for the auto-fix loop — a red PR at push time
+   is indistinguishable from drift after the fact.
+4. Only once build + tests pass on the final diff: proceed to
+   pre-PR expert review.
+
 ## Pre-PR expert review — mandatory before `gh pr create`
 
-After the branch is pushed but **before** opening the PR, run a
+After build + tests are green but **before** opening the PR, run a
 second expert pass on the actual diff. The Step 4 synthesis
 reviewed the plan; this step reviews the code. They catch
 different things — protocol drift, broken tests, overlong files,

.changeset/triage-build-test-gate.md

@@ -0,0 +1,4 @@
+---
+---
+
+Triage routine now runs a mandatory pre-PR build + test gate (npm run precommit or equivalent) before expert review, with 2 build→fix iterations. Also elevates bullet-label boundary framing ("**X gaps**") from CI warning to hard error in the current-context lint, and adds an explicit rule to the context-refresh prompt.

server/tests/unit/rules-loader.test.ts

@@ -38,7 +38,6 @@ describe('Rules Loader', () => {
     expect(rules).toContain('## Partner Directory');
 
     // Knowledge
-    expect(rules).toContain('## AdCP Agent Types');
     expect(rules).toContain('## Prebid Expertise');
     expect(rules).toContain('## Trusted Match Protocol (TMP)');