fix: follow-ups to #306 (baseline cost, CRLF, gate wording)

[MendixMau]

Follow-ups from an independent review of the merged #306 (change log, squash 3c34389). Refs #306.

1. Baseline context cost

Problem. #306 made skills/change-intake.md (1,804 words) a baseline file for lead and ba at Stages 2–7, and moved nothing to on-demand to make room. bin/render-routing.sh's budget header says a new baseline row has to be paid for elsewhere.

Fix. The file is split in two:

• skills/change-intake.md stays baseline, cut to a 299-word core. It keeps the trigger, the row columns and statuses, class → entry artifact, size → ask, a short Owners section (so the existing "change-intake.md → Owners" references in the runbook, agent-roles.md and both agent stubs still resolve), and Done when. It points to the companion file.
• New on-demand skills/change-intake-detail.md (routing row: lead, ba, architect; Stages 2–7; group spine). It holds the why-table, the wrong-vs-right example, the full Step 1–5 text (class table with the "What changed" column, the size table, the ask table with status paths, Apply, the full Owners table, Prove), the mechanical-check section, and the per-line checks for Done when.

Removed vs moved. Everything from the old file after its header is in the detail file verbatim. The only old lines not carried over word for word are:

• the old title, **Purpose:** and **Owns:** header lines. They were rewritten for each file's role. Ownership still sits with change-intake.md, and the detail header says so.
• the heading ## Done when (denominators). Its five bullets moved to the detail file as "Checking each Done-when line", and the core has a condensed version.
• the Gate-line bullet `Change log:` reads `0 of M … not yet applied`. Gate-check never prints that string, so it is replaced on purpose (item 3).

No rule, denominator or failure example was dropped.

bin/render-routing.sh --check (LC_ALL=C):

• Baseline tier: 79,997 → 78,492 words (budget 80,000).
• Stage 2/3 lead: 34,414 → 32,909.
• ba at Stages 2/3: 29,042 → 27,537.

The budget value is not lowered here, because render-routing.sh is outside this PR's file list. The ratchet rule suggests lowering it at integration.

2. CRLF PROJECT.md

Problem. mxtk_change_log_has_section used grep '^## +change log *$', which never matches ## Change log\r. So on a CRLF register:

• gate-check printed "no '## Change log' section"
• sync-project.sh step 2b1a appended a second empty section
• the scan counted 0 rows

Fix (in bin/lib/change-log.sh only, which covers both callers): both readers drop a trailing \r before matching, as gate-check already does for the tables it reads. The heading test is now a single awk call rather than tr | grep -q, because under the callers' pipefail an early-exiting grep -q can SIGPIPE the producer and turn a hit into a miss.

Fixture. New case T9 in tests/wave2/test-change-log.sh uses a real init scaffold with 4 rows, converted to CRLF. It checks that:

• the section is found
• the scan output matches the LF copy byte for byte, with no \r leaking into it
• the gate line matches the LF copy
• sync leaves exactly one heading and does not report an append

Against master's change-log.sh, 5 of these cases FAIL. The CRLF scan reads SUMMARY 0… and sync writes 2 headings.

The fixture also drops pipefail. Under it, an early-matching printf "$OUT" | grep -q could SIGPIPE, and T7 "sync says what it appended" failed intermittently on one run. No assertion in the fixture reads an upstream exit code.

3. Done-when wording vs the gate line

The skill expected 0 of M … not yet applied, but when nothing is pending gate-check prints PASS — N change request(s); every one applied with evidence, proven, or rejected (or PASS — 0 change requests logged). The skill text was changed and gate-check was not. The core now says the line "reads PASS", and the detail file quotes both PASS forms. The only change to gate-check.sh is its re-rendered stage map, which now lists change-intake-detail.md.

Not in this PR: exec.sh --cr position

project-bin/exec.sh is owned by another PR. Its --cr flag only works as the first argument, although the usage line says [--commit-first] [--cr CR-NN]. A patch with a leading-flag loop that also refuses a script name starting with -- is handed to the lead for that PR.

Checks run

• bash tests/wave2/test-change-log.sh bin/gate-check.sh: 47 passed, 0 failed. The same fixture on master: 43 passed, 5 failed (exactly the CRLF cases).
• test-baseline-pack.sh: 16 ok. test-routing-tier.sh: 12 ok. test-bug12-sync.sh: 61 pass.
• test-render-routing.sh: 5/5. One earlier run showed 4/5 because of a flaky UNROUTED report that also happens on master. Cause: routing_unrouted_skills in bin/lib/skill-routing.sh runs printf | grep -qxF under pipefail and gets SIGPIPE. That file is outside this PR.
• bin/render-routing.sh --check: in sync (same flake possible).
• bin/check-scripts.sh, bin/check-portability.sh, bin/check-no-private-citations.sh, leak guard (LEAKGUARD_ALLOW_NO_DENYLIST=1), bin/check-pr-discipline.sh origin/master: all clean.

Field run

Still owed. The CRLF fix and the split were exercised only on real init-project.sh scaffolds (golden input, as T9 builds it), not on a live project's register.

Risks

• Existing projects: no PROJECT.md behaviour change for LF files. A CRLF file that already got a duplicate section from sync keeps it, and that needs a manual delete.
• An agent that reads only the core has less detail. Each core section points to the detail file, and the routing row fires on "applying or proving a CR".

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw

---

Generated by Claude Code

[MendixMau]

Landed on master as part of the integration squash #329 (commit 31e90ae), after independent review — its Review: line is in that commit message. Closing as superseded.

---

Generated by Claude Code