[MendixMau]

Direct lane

What changed and why: #687 asks for a time budget that works like the word budget. Without one, each PR adds a per-exec or per-row step and nobody adds up the total; #686 measured an average of 19 min per row. This PR adds four things:

1. PR template: the Runtime row asks for a Time cost: line. It covers granularity (per exec, row, module, stage, project, or none), the measured seconds and the machine, whether the step can force rework (y/n), and the offset.
2. docs/RELEASE-PROCESS.md §4a: a new Per-row budget: section. It has the 2026-11-06 baseline: 128 process starts per clean MDL exec; in the Research: Stage 5 per-row cost (avg 19 min, up to 41 min per build-plan row), where the time goes and what the toolkit should prescribe instead #686 field run, about 18 s per exec and 19 min per row. Its rule: a raise needs an offset or an explicit maintainer OK, and that is recorded in the Review: line. Prefer per-module or per-stage steps (perf(exec): take the mxcli report score at module close, not after every exec #615). The reviewer re-measures. RM-02 (process/roadmap.md) is cited for the per-release measurement.
3. CLAUDE.md "Merging a PR": one sentence: the reviewer re-measures time cost for PRs that touch hot-path files.
4. tests/wave2/test-exec-process-budget.sh with the baseline in tests/wave2/fixtures/exec-process-budget.txt. The fixture runs one clean MDL exec on a minimal scaffold under env -i. PATH is closed: shims log each start, then exec the real tool. mxcli, mxbuild, java, sqlite3 and pgrep are logging stubs. An exported command_not_found_handle counts any tool that is not on the list. The fixture FAILs above baseline + 5%. Lowering the baseline is free; raising it shows up in the diff. Background refresh is forced synchronous (MXTK_SYNC_REFRESH=1), because racing it made the count vary. Lint output is the verbatim v0.28.0 capture. run-all.sh picks the fixture up through its usage: line (subject exec.sh).

Skipped: the advisory diff check (item 5). bin/check-pr-discipline.sh does not read the PR body.

Resolution, read this. 5% of 128 is about 6 starts. One extra process, or a whole second lint-gate pass (+6 → 134), still passes, with a "rose by N" note. A step of 7 or more starts fails. Process count is a proxy: one mxbuild is one start but takes minutes. That is why the reviewer still re-measures seconds.

Baseline may need a re-measure after perf/module-close-proofs-686 merges, because it changes exec.sh's lane text. The baseline was measured on origin/master ee12a615.

Impact

Row	Answer
Process	No new blocking step for projects. One new CI fixture can turn a PR red when exec.sh's process starts rise more than 5%. The PR body now owes a Time cost: line for new steps.
Dependencies	Read only, not changed: project-bin/exec.sh and the rest of project-bin, copied into a scratch scaffold. The baseline depends on exec.sh; see the re-measure note for the sibling #686 PR. Reads the existing capture fixtures/lint-conv020-v028/.
Context cost	none: no routed or always-loaded file touched (render-routing.sh --check passes)
Behaviour	NOT RUN: process docs and a template that review agents read. No agent-stub or skill text changed.
Runtime	Projects: none. CI: about 1.5 s for the new fixture.

Time cost: per PR (CI) · ~1.5 s on Linux x86-64 sandbox · can force rework: y (a hot-path PR over budget) · offset: none needed, not on any project path

Field evidence: measured against origin/master exec.sh: 128 starts, the same over 6 runs, 1.1–1.5 s each. The numbers come from #686's field run (163 execs, 17 row agents). Fail demo: a dummy step of 7 date calls added to exec.sh after the lint-gate call gives 135 > 134, FAIL, rc=1, with a per-tool breakdown. Reverted, it gives 128, PASS. Linux only: the fixture SKIPs elsewhere. The count is not yet confirmed on the GH ubuntu runner, so CI on this PR is the first cross-machine check.

Checks run: the new fixture (PASS=4), check-scripts, check-portability (clean after portability-ok markers on the Linux-only shim names), the leak guard, check-pr-discipline, and render-routing --check. No full suites.

• No client data anywhere in the diff
• Size cap: 6 files, well under 400 lines
• CHANGELOG line, credited "— Research: Stage 5 per-row cost (avg 19 min, up to 41 min per build-plan row), where the time goes and what the toolkit should prescribe instead #686/Process-cost guard: a PR that adds a check, gate, obligation or test step must state its time cost per build row, under a budget #687 speed audit"

Review: pending independent review

---

Generated by Claude Code