refactor(acm): perfect dao-shu guidance — unconstrained Composer synthesis

[KorenKrita]

Summary

Unconstrained always-on ACM guidance: everything retained from #6–#9 when token budget is not the limiting factor (~9304 chars vs ~8500 lean ultimate).

What this adds over #8/#9

Addition	Why
How judgment layers fit together	One integrator paragraph: smells → moments → criteria → failure shapes (道术 wiring without a second table)
Best combined opening	keep live what NEXT needs + recoverable handoff + NEXT executable
Full invariant tie-back	against this invariant everywhere it matters
Full lightest-transition bridge	Complete situation wording
Dual skill pointers	Rebase inline + end-of-CORE (full skill name)
Cold start + anchor gravity	In target-selection line and pressure paragraph
Managed-chain carve-out	Lightweight reply explicitly outside managed-chain workflow
Full post-rebase execution	without rereading archived summaries restored

Unchanged good choices

• Single Recognizable moments table (no Judgment duplicate)
• 4 smells; NEXT stalls only in Failure shapes
• Fold/Rebase criteria (soft naming)
• Why-annotated examples + travel self-check
• Nudge ↔ CORE vocabulary chain
• handoff working set cue

Structure

Causal keep-live opening
Working-set invariant
How judgment layers fit together   ← new integrator
Vocabulary
ACM preflight (cost curve)
Decision smells (4)
Recognizable moments + lightest bridge
Fold criteria + self-check
Rebase criteria + inline skill pointer
Failure shapes
Handoff + Why examples
Skill pointer (full list)

vs lean ultimates

	#9	#8	Perfect
CORE chars	8496	8732	9304
Judgment layers integrator	—	—	✅
keep-live opening	✅	partial	✅
Dual skill pointers	—	partial	✅
managed-chain carve-out	—	—	✅

Recommend for users/models where ~800 extra always-on tokens are acceptable. For strict budget, prefer #9.

Test plan

• bun run verify:acm
• Live: read layers once → smell → moment → criteria self-check → travel

Supersedes

Optional merge target; lean production may still prefer #9. Closes the design exploration from #2–#8.

Worktree: ../pi-context-perfect

[KorenKrita]

Superseded by #12 (chosen merge target for ACM dao-shu guidance).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