chore(context): per-role on-demand token in the routing agents column (stub diet T1) — stacked on #378

[MendixMau]

Stacked on #378 (chore/stub-diet-runbook). This branch contains #378's two commits. Review only the top commit (a406512) until #378 merges; then this diff shrinks to that commit.

Direct lane

What changed and why (one paragraph):
This is item T1 of the stub-diet context-cost research (r-stub-diet). The tier column in bin/lib/skill-routing.tsv applies to the whole row. So a row that one role needs every session, and another role needs once per module, had to be baseline for both. Example: the mdl helper never closes a module, pushes or writes a gate script, yet it read module-review, ui-loop, teamserver-alignment, learned-css-that-never-applied and skills-over-scripts every session.

The fix is a new agents-cell token, <role>? (for example mdl?, or all,mdl?). It makes a baseline row on demand for that one role:

• Stub: the row stays in the role's stub, among the on-demand rows below the baseline rows, with the trigger only. It stays findable (CLAUDE.md skill rule 8).
• Word count: the row leaves that role's always-read set and word count.
• Other roles: every other role the cell names keeps the row at baseline.

Where it is handled:

• _routing_role_tier is new; _routing_reaches and routing_baseline_pack respect the token.
• The agent:<x> view places the row by its per-role tier.
• The Role(s) column renders review (mdl: on demand) and every role except mdl, review (on demand for those). The CLAUDE.local.md and stub preambles each gain one clause saying what "on demand" means.
• The BAD_AGENT audit checks the name part of a ? token. It refuses ? on a non-baseline row and mdl,mdl? (both forms of one role).

The token is applied as follows:

• mdl: module-review, ui-loop, teamserver-alignment, learned-css-that-never-applied, skills-over-scripts.
• review: teamserver-alignment, skills-over-scripts.

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

	master d16f578	after #378 (A)	after this PR (A+B)
Baseline total (budget 80,000)	75,028	73,233	73,233 (unchanged: rows stay baseline for other roles)
Stage 5 mdl	56,236	43,352	30,400 (target ~31k)
Stage 5 review	44,818	31,934	30,018
Stage 5 gate / test	34,629 / 33,689	21,745 / 20,805	unchanged
Stage 6 mdl / review	32,983 / 39,723	19,875 / 26,615	9,969 / 24,699
Stage P–3 mdl / review	21,106	7,694–8,603	6,642–7,551

T4 (verbatim narrative moves, about 1.5k) is a separate PR. T5 (detection-gaps register) is untouched while it waits for a maintainer ruling.

Mechanical no-loss check: tests/bench/trim-noloss.sh chore/stub-diet-runbook HEAD (the bench added in #378). Result: 6 lines found elsewhere, 17 routing rows still routed, 0 left. Every removed stub line is a routing row re-rendered in its on-demand form.

Owed before merge: behaviour A/B (maintainer's session, measurement plan from r-stub-diet, process/preflight-skill-baseline-2026-08-31.md method):

• Fresh Sonnet drafting agents, 5 reps per arm. One arm gets the master stub and skills, the other gets A+B. Two tasks: one page task, and one microflow task with a loop.

• Score each run on:

  • shell violations (check-page-shell.sh)
  • page-fidelity.js --no-log
  • the report block with its denominators
  • write mode per operation
  • whether the microflow-preflight checklist is posted
  • whether a seeded STOP operation is refused

  Pass if no metric is worse than n=5 noise.

• Retrieval probe (skill rule 8), specific to this PR: seed each re-tiered trigger, for example "the style change did nothing" (learned-css), "close this module" (module-review), "the page looks wrong" (ui-loop) and "push to Team Server" (teamserver-alignment). The mdl helper must open the right file in 5 of 5 reps.

• Start tokens per helper from bin/context-audit.sh on the next field build.

Field evidence: none yet.

• No client data anywhere in the diff
• Size cap: the top commit changes 13 files and about 196 lines. Excluding tests and CHANGELOG it is 11 files, and 7 of those are generated re-renders (README, ROUTING, agent stubs).
• Test tier: T0, plus T1 scoped fixtures, plus render-routing --check. All green: test-routing-tier 26/0 (new cases T6–T8; 9 FAIL against chore(context): runbook slice for helpers; runbook drops its copy of the baseline table (stub diet T2+T3) #378's lib as the negative control), test-baseline-pack 16/0, test-render-routing 5/0, test-pipefail-grep-q 3/0, test-runbook-slice 21/0, test-ritual-slim 17/0, test-bug12-sync 61/0. Also clean: bin/check-skill-routing.sh, check-scripts, check-portability, the leak guard, private citations and pr-discipline. T2 field run: none.
• No new instrument; renderer/table change only, bash 3.2 / POSIX awk
• CHANGELOG.md line appended, crediting r-stub-diet
• No bug entries

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw

---

Generated by Claude Code

[MendixMau]

Landed on master in #385 (squash 868b3cb), after its own independent review on the night integration branch. Closing this PR.

---

Generated by Claude Code