Redesign ACM around working-set judgment

[KorenKrita]

Summary

• replace prescriptive ACM tool choreography with a leading-word working-set doctrine: working set, active uncertainty, recoverability, boundary, handoff/cold start, summary debt, and anchor gravity
• keep CORE as the way (道) while progressively disclosing technique (术) through tool contracts and the advanced context-management Skill
• remove checkpoint/travel suffix-based cue classification and use state-neutral runtime guidance
• add a non-CI, real-tool-call behavior eval with paraphrased scenario families and an independent judge
• document the ubiquitous language in CONTEXT.md and align README, AGENTS, implementation notes, generated guidance, context nudges, and tests

Deterministic verification

• bun run verify:acm
  • root suite: 91 tests, 506 assertions
  • TypeScript typecheck: pass
  • exact-host fixture: 23 tests, 144 assertions
• npm pack --dry-run --json: package includes CONTEXT.md, eval resources, scripts, skills, and runtime source

Open-ended behavior evaluation

Candidate: local-openai/mimo-v2.5
Judge: local-responses/gpt-5.4

• full scenario-family run: 10/12 passed (83.3%), above the 80% threshold, with every required family represented
• strict focused runs:
  • recoverability: 2/2
  • cold-start handoff: 2/2
  • travel isolation: 2/2

The eval is intentionally outside verify:acm because it spends model calls and measures stochastic behavior rather than deterministic build correctness.

Review

A full diff review and focused re-review completed with PASS and no unresolved P0-P2 findings. The re-review confirmed the duplicate Skill router text and Call/Run wording drift were fixed, and checked eval tool safety, guidance ownership, packaging, and test contracts.

[KorenKrita]

Superseded by the redesign now on main: CORE.md teaches working-set judgment with the six moves (save/orient/fold/rebase/rehydrate/fork), suffix-driven cue branching was removed from the runtime, and prompt metadata is generated from the canonical sources. The leading-words direction this PR explored shaped that final form.