feat(spec): Model and Reasoning routing hints (SPEC-002)

[joshuaboys]

Adds optional-but-preferred Model: and Reasoning: fields so a plan can say
which AI model should execute a work item and how hard it should think.
Implements SPEC-002 under decision D-047.

The spec

• Work items take - **Model:** (free-form identifier, e.g. claude-opus-5)
  and - **Reasoning:** (low | medium | high | max, case-insensitive).
• Modules can carry Model and Reasoning metadata-table columns as a
  module-wide default. An item field overrides the column, mirroring the
  existing Packages: inheritance.
• Neither field is required, so no existing plan gains a warning by omission.
  Templates, aps-rules.md, the planning-skill reference, and the prompts
  steer authors to fill them in.

CLI behaviour

Landed in lockstep across Rust, bash, and PowerShell per D-039.

• W023 warns only when a present Reasoning value falls outside the
  vocabulary. Placeholder prose is skipped, and Model is never
  vocabulary-checked because identifiers are harness-specific.
• aps next adds a Model: … | Reasoning: … line with the effective
  values, omitted entirely when neither is set.
• aps export carries model and reasoning on modules and work items,
  null when unset.

Verification

• cargo test (197 pass), cargo fmt --check, cargo clippy --locked --all-targets -- -D warnings clean.
• ./test/run.sh passes, including a new test over test/fixtures/routing/
  covering inheritance, item-level override, W023 on both the table column and
  the item field, the next output, and export.
• ./test/cli-parity.sh reports bash = Rust = PowerShell across all 11
  fixtures; test/ps-parity.ps1 passes. Both ran against a fetched pwsh.
• aps lint plans and markdownlint clean.

Design note

Reasoning has four levels; harnesses map max onto their own top tier
(xhigh, max, and so on). Making Model a validated enum was rejected
because a model list rots faster than the spec does.

🤖 Generated with Claude Code