refactor(acm): optimal dao-shu guidance — Composer synthesis of #4+#5

[KorenKrita]

Summary

Optimal merge of #4 Composer v5 and #5 Grok synthesis, against the original goal: models understand and self-manage context via invariant + criteria, not ritual state machines.

Base: PR #5 (leaner, less duplication, nudge↔CORE vocabulary chain)
Cherry-picks from PR #4 (stronger 道 bridges without restoring the duplicate Judgment table)

道

• Working-set invariant — single north star
• Causal opening — process accumulates naturally; ACM keeps growth recoverable (from refactor(acm): dao-shu guidance v5 — Composer refined from Grok (#2+#3) #4)
• Cost-curve preflight — why checkpoint before managed work
• Recognizable moments + lightest transition bridge tied back to invariant (from refactor(acm): dao-shu guidance v5 — Composer refined from Grok (#2+#3) #4, no second table)

术（识别）

• Decision smells — 4 quick scans against invariant (NEXT stalls stays in Failure shapes only)
• Failure shapes — post-fold consequence teaching
• Fold gate + one-line travel self-check
• Why: annotated handoff examples

术（触点）

• Nudges name decision smells, recognizable moment, fold gate, working-set invariant
• Positive tool cues; mixed-batch rejection at runtime only
• Skill router at rebase failure and end of CORE

vs #4 and #5

Piece	#4	#5	This PR
Invariant spine	partial	✅	✅
Causal opening	✅	—	✅
Single situation table	two tables	✅	✅ + judgment bridge
NEXT stalls in smells	✅	—	— (failure shapes only)
Nudge↔CORE section names	partial	✅	✅
Inline rebase skill pointer	✅	—	✅
CORE length	~8961	~8500	8763

Test plan

• bun run verify:acm
• Live: smells → invariant check → fold gate self-check → travel; pressure nudge → rebase check not ritual travel

Supersedes

Recommend merging this and closing #2, #3, #4, #5.

Worktree: ../pi-context-composer-v6

[KorenKrita]

Superseded by #12 (chosen merge target for ACM dao-shu guidance). Closing this alternative.