docs(harness): define the minimal harness interface + progressive-enhancement tiers

[slowdini]

Part of #113 ("Instant magic"). The Epic 1 "instant magic" deliverable — the conservative slice: a documented minimal interface, not yet data-driven descriptors (that's #E).

Why

#113's core goal: stop forever playing catchup on harness compatibility. Today a new harness must be wired across two dispatch mechanisms and a ~20-method HarnessAdapter trait, audited via docs/harness-parity.md. Once the in-session mechanism is gone (#B), the "two parity axes" framing collapses to one, and we can express compatibility as a minimal interface every harness must satisfy + optional capabilities that raise fidelity — and document the known mappings instead of treating every method as required.

Honest floor (must be reflected in the framing): three concerns are genuinely per-harness code, not documentable mappings — transcript parsing (claude -p stream-json vs Codex item.completed), native skill staging/rendering (.claude/skills Skill-tool list vs .agents/skills ## Skills vs .opencode/skills <available_skills> XML), and write-guard hooks. So "any harness via docs alone" applies to the baseline, with fidelity as opt-in code.

Scope

• Rewrite docs/harness-parity.md around one dispatch mechanism + a baseline/enhancement contract.
• Define the minimum operating list (zero harness-specific code): (1) a headless exec command template (invoke with a prompt, capture stdout/stderr); (2) the agent writes outputs/final-message.md, or --no-stage inlines the skill into the prompt. This already yields llm_judge grading + the detect-stray-writes post-pass.
• Define progressive enhancements, each a documented adapter capability with a default/fallback:
  • transcript parser → unlocks transcript_check + token/cost capture;
  • native skill staging dir + skills block (vs --no-stage inlining);
  • model flag;
  • write-guard hook (vs detect-stray-writes fallback).
• Audit the HarnessAdapter trait: annotate each method baseline-required vs optional-with-default, and add default impls so a baseline harness need not implement the optional ones.
• Ship the "known mappings" table Instant magic (what a clever name!) #113 envisioned: harness → exec template, events filename, skills dir, skill-block format.
• README "Harnesses" intro: point at the reframed doc.

Files

docs/harness-parity.md (rewrite), src/adapters/harness.rs (trait doc comments + default impls), README.md ("Harnesses" intro).

Acceptance criteria

• The doc describes a single mechanism and a clear baseline-vs-enhancement contract.
• The trait visibly marks baseline-required vs optional-with-default methods.
• A reader could wire a baseline harness from docs + --no-stage without writing a transcript parser.

Dependencies

Depends on #B; benefits from #C landing first.

---

Filed from the #113 feasibility/breakdown plan.