Migrate semantic interpreter to task_complexity_scale (INTENT §4 v3 compliance)

[CoreyRDean]

Background

INTENT.md §4 (architectural properties, v2) added "Cost transparency": every Claude session clauck spawns derives sizing from a single shared formula in lib/sizing.py keyed on a task_complexity_scale float. The scope statement named the semantic interpreter as a known compliance gap that must be closed before the v3 contract amendment:

Scope of compliance at v2 adoption: doctor, scheduler, CLI surfaces (clauck size, clauck validate, clauck inspect), and all marketplace jobs are fully compliant. The natural-language semantic interpreter (clauck <anything>, clauck work <text>) still emits its own stage-2 execution params directly — a known compliance gap tracked for follow-up.

What's non-compliant

_build_interpreter_prompt in lib/clauck (search: def _build_interpreter_prompt) instructs the classifier to emit:

{
  "command_type": "semantic",
  "interpretation": "...",
  "enhanced_prompt": "...",
  "exec_model": "haiku|sonnet|opus",
  "exec_effort": "low|medium|high",
  "exec_max_turns": <int>,
  "exec_max_budget_usd": <float>
}

The last four fields are exactly the literal-anchoring failure mode that drove PR #101 for doctor: classifiers under load grab literal dollar values from the prompt instead of computing, producing the $0.25 budget-exceeded cluster that broke doctor in practice. Every clauck <anything> invocation rides that same risk today.

The docstring at _build_interpreter_prompt acknowledges the gap with a pointer at INTENT.md §4, but that doesn't close it.

What the fix looks like (follows the doctor migration pattern from PR #101)

1. Classifier output shrinks to scale-only: emit {command_type, interpretation, enhanced_prompt, task_complexity_scale} — drop the four exec fields entirely.
2. cmd_semantic computes params via sizing.compute_sizing using the emitted scale, same way cmd_doctor does today.
3. Update the prompt's "Sizing the semantic execution" section with scale anchors matching the job-creation guidance already in that file (0.10 trivial / 0.25 focused / 0.40 multi-source / 0.55 investigation / 0.75 DAG / 0.95 architectural).
4. Keep the job-creation guidance as-is — it already correctly directs Claude to emit complexity: in new job frontmatter, not raw params.

Doctor's migration is the reference: see _build_doctor_interpreter_prompt + the cmd_doctor routing block around the sizing.compute_sizing call.

Why this is tracked vs. just done in PR #128

PR #128's scope was "make the plugin the canonical install path." The semantic-interpreter migration is its own refactor with its own review surface — crosses the cmd_semantic execution path, changes the routing JSON schema, and touches the only remaining consumer of the scale-less flow. Bundling would have ballooned that PR; tracking as follow-up was the right call.

Acceptance

• _build_interpreter_prompt emits task_complexity_scale only for the semantic branch; no model/effort/turns/budget fields.
• cmd_semantic derives all four params via sizing.compute_sizing (including MCP auto-promote for non-strict_mcp_config runs).
• Invocation smoke test: a trivial clauck "list my jobs" stays cheap; a clauck "audit every job for cost overruns" routes to sonnet/high with a budget the formula produces, not a classifier-guessed literal.
• Update INTENT.md §4 scope statement: remove the "known compliance gap" carveout once landed.
• Optional: amend INTENT.md version to v3 when the last scope-statement compliance gap closes.

References

• INTENT.md §4 — Cost transparency architectural property
• PR feat(sizing): complexity-scale-based derivation for all Claude sessions #101 — doctor interpreter migration (the template for this)
• lib/sizing.py — the shared formula
• lib/clauck — _build_interpreter_prompt (classifier) + cmd_semantic (stage-2 routing)