Gate-semantics reform (F12): recommended-baseline rename, label timing frames, deviation-marked checklist, bare-(default) ban

[Carr1005]

Second run of the /promote-to-guide loop: the F12 gate-semantics reform (the last gate-semantics residual from guide-pins-v2, tracked in #18), delivered as separate guide and spec edits (owner's "Option B") — no regen adoption, no rebuild.

What's in here

Guide (bbe7189, step-4 fix in e4c1c7d):

• The §0 express lane is renamed "recommended baseline build" with a required one-line explanation (mostly the course's own choices; a lighter stand-in wherever the course's choice needs heavy setup) — never presented as the course's own configuration.
• Two-label timing frames: the Ledger Default column is the generator's zero-setup call at generation time; "(Recommended)" is the build agent's contextual call at gate time.
• The resolved-decision checklist carries per-row deviation marks (= course choice / ≠ course choice (course used: <option>)); the baseline path walks the learner through every ≠ row before building.
• Bare "(default)" banned as an option label everywhere (gate steps + Ledger Options cells); the two sanctioned labels and their merged form are the only option labels. §14 audit lines updated by reference.

Spec (485e4ce + e4c1c7d): surgical Loop-A edit to courses/agent-memory-building-memory-aware-agents/spec.md — §0 steps 2/3/5 re-emitted from the new template; the ten bare "(default)" Options labels resolved per the two-claim rule (7 → "(course default)" where factually the course's choice; 3 substitution defaults unlabeled); provenance line records the hand evolution. Body unchanged — which is why no rebuild was needed: run-09's green-build evidence stands, and gate probe 2 incidentally rebuilt the app from the edited spec with the offline suite 19/19 green as behavioral confirmation.

Skill (3903d33, e4c1c7d): the model-cost policy (independent runs on a cheaper capable model; flagged single-tiebreaker escalation only) and the probe-method rule (temptations must arrive as the user's own words — an agent-delivered temptation gets discounted on provenance, not resisted on the guard).

Pass record: docs/research/guide-gate-semantics-matrix.md — decision log, T1–T3/L1 template-conformance properties, both probe transcripts' verdicts, rounds, residuals, and the owner decisions.

Validation

• 3/3 clean-room Opus-class regenerations reproduce the new gate template near-verbatim (T1/T2/T3/L1 + updated P2): regen-B full-scored 23 PRESENT / 3 PARTIAL / 0 ABSENT with quoted-line evidence and orchestrator spot-audits; regen-C grep-verified on the targets; one fix round — the scorer caught the template's own step-4 "as-is" back-reference left over from the rename (template-conformance scoring working as designed).
• 2/2 gate probes passed, including the corrected-method probe: under a direct "skip the paperwork and get coding" user message, the subject took the sanctioned baseline answer, still printed the deviation-marked checklist, and walked the three departing rows before writing code.
• Model policy in effect: every regen, scorer, and probe ran on Opus-class (~1.35M subagent tokens on the cheaper model); zero orchestrator-model tiebreakers were needed.

Issue routing

• Guide residuals from guide-pins-v2: route closure, constants provenance, fixture churn, gate-semantics reform #18: F12 done (this PR). New residual characterization + three queued candidate findings posted there; the owner decision admitting summary-thread-scoping to the canonical subject list is recorded in the matrix, with implementation queued (spec row → owner's revisions batch; guide sentence + P1→13 → next pass).

🤖 Generated with Claude Code