tech-debt: main is red — code-host-routing AC4 audit flags assess-cost/SKILL.md ("merged PRs ... from the PM tool / code host")

[rucka]

Story Statement

As a maintainer merging any open PR
I want origin/main to be green on the code-host-routing.test.ts AC4 full-catalog audit
So that every PR's merge-ref CI reflects that PR's own quality instead of inheriting a failure from main

Where: packages/knowledge-hub/dataset/.skills/capability/assess-cost/SKILL.md (Step 8 wording) vs. the audit regex in packages/knowledge-hub/src/conformance/code-host-routing.test.ts (PR_FROM_PM_TOOL).

Epic Context

Parent Epic: none — cross-cutting CI/gate regression, collision between two independently merged stories.
Status: Todo
Priority: P0 (Must-Have) — main is red, so every open PR's pull_request merge ref is red.

Status Workflow

• Todo: Story is created but not yet refined
• Refined: Story is detailed, estimated, and ready for development
• In Progress: Story is actively being developed
• Done: Story delivered and accepted

User Value

User Benefit: a red check on a PR means something about that PR again.
Business Impact: unblocks the merge gate for the whole open-PR set; removes the "is this mine or main's?" triage cost from every review.
Visible UI Value: n/a (CI) — demonstrable as a green CI check on main and on any rebased PR.

Rough Sizing

Story Points: XS(1)
Confidence: High
Reasoning: one sentence to reword (plus its .claude/skills/ mirror), or one tempering clause in the regex. The decision of WHICH is the only open question.

Initial Scope

Likely In Scope

• Resolve the failure: either reword assess-cost/SKILL.md Step 8 so the PR read is routed explicitly, or temper PR_FROM_PM_TOOL for the the PM tool / code host (both-sides-named) shape.
• Keep both copies in step (dataset source + .claude/skills/pair-capability-assess-cost/SKILL.md) per the mirror-equality guard.

Likely Out of Scope

• Any other change to assess-cost report mode (assess-cost report mode — deploy/billing telemetry integration for deploy-match #399 territory).
• Broadening or narrowing the audit beyond this shape.

Open Questions

• Is from the PM tool / code host a genuine conflation (fix the prose) or a legitimate both-sides-named read the audit should tolerate (fix the regex)? The regex already tempers on code[- ]host appearing INSIDE the gap; here it appears after the PM tool, so the temper does not fire.

Acceptance Criteria

AC1 — the audit is green on main
Given origin/main at or after the fix
When pnpm --filter @pair/knowledge-hub test runs
Then code-host-routing.test.ts → "AC4 — no skill doc in the catalog conflates a PR with the PM tool (full-catalog grep)" passes with an empty offender list.

AC2 — the routing intent is preserved, not silenced
Given whichever side is changed
When the fix lands
Then assess-cost Step 8 still states that the merged-PR list is read from the code host and the item state from the PM tool, and the audit still fails on a genuine conflation (a regression case pins whichever shape was tempered).

AC3 — both copies stay in step
Given the dataset SKILL.md is edited
When the mirror guards run
Then .claude/skills/pair-capability-assess-cost/SKILL.md matches, and skills:conformance / docs:staleness stay green.

Definition of Done Expectations

• Functionality implemented and working (CI green on main)
• Automated tests written and passing (regression case for the tempered/reworded shape)
• Code reviewed and merged
• Documentation updated (only if the audit's documented rule changes)

Dependencies

Story Dependencies: #236 (introduced the audit, landed as ADR-018), #388/#399 (introduced the assess-cost report-mode wording).
Epic Dependencies: none.

Notes

Found during the independent review of PR #411 (story #407) — not a defect of that PR, whose own head is green on every gate locally and merges cleanly into main.

Failure, verbatim (build job → @pair/knowledge-hub#test):

FAIL src/conformance/code-host-routing.test.ts > code-host / PM-tool split — no skill assumes the two coincide (#236, AC4 audit)
  > AC4 — no skill doc in the catalog conflates a PR with the PM tool (full-catalog grep)
AssertionError: expected [ 'capability/assess-cost/SKILL.md' ] to deeply equal []

Offending sentence — capability/assess-cost/SKILL.md, Step 8:

Read the window's merged PRs — closed PRs whose merge date falls inside the window — from the PM tool / code host

First red commit: f1e581b7 (#388). Still red at 768e8a25 (#391) and 8f8956bd (#408) — i.e. three merges have landed on a red main.

Why nothing caught it pre-merge: the audit is a full-catalog grep, so it is only tripped by the combination of #236's regex and #388's prose — neither PR's own branch contained both. File-level disjointness analysis does not catch it; this is the same behavioural-collision class PR #411's body flags for its own #406 coupling.