feat(os): model & effort routing policy as data + ADR-0017 (M16 16.1)

[danielPoloWork]

Closes #252 (M16 16.1). Plan: .issues/M16-model-effort-routing-milestone.md (#256). ADR: ADR-0017.

What

The eighth OS spec, orchestrator/os/routing/ — "which model, how much effort?" as
policy-as-data instead of a manual per-issue judgment.

• routing.yaml + _schema.md — work signals (tracker labels + the asserted
  sets-pattern / decision-heavy flags) resolve by monotonic escalation: start at the
  cheapest floor (fast/low), every matched rule raises to at least its
  min_tier/min_effort, final = max. Rules never lower — the same only-add-never-relax
  principle as domain overlays; deterministic and order-independent.
• Tiers, not model names. Rules speak fast/standard/frontier-reasoning; names live
  only in the dated catalog: (today, Claude Code: Sonnet 5 / Opus 4.8 / Fable 5) with
  per-host effort_aliases ("ultracode" → max). Model churn = a catalog edit + refreshed
  as_of:, never a policy/code change. Precedent: risk-weights-as-data (feat(risk): risk-model weights as per-domain data in risk.yaml (M6 6.2 / F3) #79), ADR-0009.
• The tier call ships as a M14: examples: blocks — judgment-laden guidance as validated knowledge-as-data #224 worked-example decision surface (6 cases seeded from the
  maintainer-reviewed M15 assessment), registered in eados_lint EXAMPLE_FILES — the only code
  touch (one tuple entry).
• ADR-0017 records the posture: advisory-first, human keeps final model authority
  (RFC-0001); auto-application only for delegated subagent work (lands with 16.4/[16.4] Auto-apply routing where the host allows: delegated subagent work carries model + effort #255).
• Docs: os/README.md spec table, AGENTS.md §4 spec list, ADR index, CHANGELOG [Unreleased].

One deliberate deviation from the issue letter

#252 sketched a separate models.yaml; the catalog is instead a top-level catalog: key of
routing.yaml, because os-spec-completeness schema-validates only the <spec>.yaml
instance — a sibling file would get parse-checking but no schema keys. Folding it keeps the
whole policy under the strongest existing gate. Same data, better perimeter.

Verification

• eados_lint.py self-lint: OK (os-spec-completeness auto-discovered the new spec; os/**
  gate-coverage pattern already applies — no registry change needed).
• Full discovered suite: 46/46 OK (run_all.py), incl. test_examples over the new block.
• Differential parse: yamlmini.load_yaml vs PyYAML safe_load on routing.yaml →
  structurally identical (incl. quoted as_of staying a string in both).
• Referential integrity (rules→tiers/efforts/flags, every host maps every tier) is documented as
  _schema.md invariants; loud enforcement lands with the 16.2 evaluator ([16.2] tools/route_advice.py — pure evaluator: issue signals → recommended tier + effort + rationale #253) per the ADR.

🤖 Generated with Claude Code