feat: per-host SKILL.md generated from shared template (#55)

[earino]

Summary

• Closes v0.5: Generate per-host SKILL.md from a template #55. Each AI CLI host now reads a SKILL.md that only mentions its own tool names — no cross-host enumeration. Eliminates the qwen3.6-35b mis-dispatch class observed in samples-issue42/session-ses_235c.md where the model picked Claude Code's general-purpose (listed first in the cross-host bullet) on its OpenCode dispatch and got Unknown agent type rejected before self-correcting to general.
• SKILL.md.tmpl is the source of truth; tools/gen_skill_md.py emits SKILL-<host>.md per host (claude / codex / gemini / opencode) plus a default SKILL.md (= claude variant) so a clone-without-install user still gets a working file. Template engine is intentionally minimal — {{var}} substitution + {{#if host}}...{{/if}} / {{#unless host}}...{{/unless}} conditional blocks, no Jinja2 dependency, ~120 lines of stdlib Python.
• install.sh detects the install path (.claude/skills/... → claude, .codex/skills/... → codex, etc.) and copies the matching variant over SKILL.md. Clones to non-canonical paths default to claude.
• tests/test_skill_md_template_generation.py (16 tests) pin drift detection (--check exits non-zero if regenerated content would differ), per-host shape (each variant references ONLY its own tools; OpenCode keeps its OPENCODE_ENABLE_EXA + parallel-flake notes; Codex keeps its prose-instruction dispatch; Gemini keeps @generalist), and install.sh wiring (case-statement per host, swap references SKILL-$DETECTED_HOST.md, catch-all defaults to claude).
• Modest size reduction (~3% bytes, ~27 lines per variant) — the issue's 15% estimate over-counted because cross-host enumerations were tighter than expected. The disambiguation win is the real value, and the template exists for the next time we add a host (one row in tools/host_config.py + one branch in install.sh).

Test plan

• python tools/gen_skill_md.py --check exits 0 in a clean tree.
• pytest tests/test_skill_md_template_generation.py -v — 16/16 pass.
• pytest tests/ -q full suite — 476/476 pass (was 460 + 16 new).
• Per-host spot-check: OpenCode variant has task / general / OPENCODE_ENABLE_EXA / sst/opencode#14195; Claude variant has Task / general-purpose and no general standalone; Codex variant explains the prose-instruction dispatch shape; Gemini variant uses @generalist.
• install.sh path detection traced manually for each canonical install path (claude / codex / gemini / opencode) plus a non-canonical fallback.
• End-to-end run on a non-Claude host (deferred — Claude Code variant is the only one most cohort users see at install time; the templating system itself is exercised by the test suite).

🤖 Generated with Claude Code

[earino]

Closing per maintainer call. The right move at v0.5 is simplification (tighten SKILL.md while keeping every prescription) rather than templating-on-top-of-redundancy. The qwen failure modes we shipped fixes for in v0.4.4 were all 'prescriptiveness at ambiguity points,' not 'doc length' — and trimming the 250-line telemetry reference + 150-line debugging playbook + 9 expanded Phase 0 question shapes should help weak-model reliability (more of the load-bearing Phase 0–7 content stays inside OpenCode's tool_output cap). Templating can return as v0.6+ if simplification doesn't suffice. Branch feature/issue-55-templated-skill-md and the SKILL.md.tmpl + tools/gen_skill_md.py + per-host SKILL-<host>.md artifacts are preserved on the branch in case we want them back. New issue with the simplification plan: forthcoming.