feat(doctor): wire into the 9 per-type templates, the Windows dispatcher, and the plugin-path SKILL.md

[fujibee]

Wires doctor (default = whole install, --project/--type/--team/--redacted) into the command reference every install path deploys. doctor itself already shipped (#654); this PR is purely wiring, no behavior changes to doctor.sh.

Addresses part of #267.

Why 11 files, not 9 or 10

There are two separate install paths, each deploying its own copy of the command reference, and both needed wiring:

1. ./install.sh deploys scripts/drivers/types/<type>/template.md, one per agent type — 9 files (agmsg-app has no template; it's spawnable=no, the desktop app's own identity, not a spawnable CLI type, so it's correctly excluded, not a 10th type I missed).
2. The Claude Code plugin-marketplace install path does not run install.sh at all — it drops this repo's root SKILL.md straight into ~/.claude/plugins/cache/ as the deployed skill content (see SKILL.md's own Step 0 comment). This is a 10th file, easy to miss by only looking at install.sh's deployment tree.

Plus scripts/windows/dispatch.sh (the one dispatch/router file in the repo, confirmed via tests/test_dispatch.bats as its regression coverage) — 11th file.

Checklist — please verify all 11 are present and non-trivially changed:

• scripts/drivers/types/antigravity/template.md
• scripts/drivers/types/claude-code/template.md
• scripts/drivers/types/codex/template.md
• scripts/drivers/types/copilot/template.md
• scripts/drivers/types/cursor/template.md
• scripts/drivers/types/gemini/template.md
• scripts/drivers/types/grok-build/template.md
• scripts/drivers/types/hermes/template.md
• scripts/drivers/types/opencode/template.md
• scripts/windows/dispatch.sh
• SKILL.md (repo root — the plugin-path deployed copy, not a per-type template)

Mechanical check used while writing this: grep -c doctor <file> on all 11 before/after — every target went from 0 mentions to a non-zero count; no target already mentioned doctor beforehand (no partial/stale-reference risk).

What each file gets

The 9 templates: a bare doctor bullet in the post-join command summary, plus an argument-dispatch block. Each matches that template's own existing conventions (invocation prefix — /__SKILL_NAME__ vs $__SKILL_NAME__; hermes's natural-language "ask X" phrasing instead of a slash form) rather than one block copy-pasted into all nine — a template that doesn't match its own type's conventions can still pass CI silently.

Every dispatch block documents bare doctor (no flags) as the normal form, with --project/--type/--team/--redacted described as optional narrowing — not the headline usage — since the default scope is already the whole installation.

scripts/windows/dispatch.sh: a doctor case that passes its argv straight through, with no identity resolution. doctor needs no TEAM/AGENT, and this dispatcher's own --project/--type/--team globals serve a different purpose (identity-resolution context for other commands) than doctor's own same-named scope flags — conflating them would silently change what a Windows user's --project/--type meant depending on which subcommand followed. Two new tests in tests/test_dispatch.bats cover: no-identity-needed default scan, and flags passing through untouched by the dispatcher's own globals.

SKILL.md (root): one addition to the "Shell (any agent)" command block, written in that section's own existing prose-comment style, not the templates' numbered-dispatch style — it's a different kind of document (a distributed end-product, not a per-type instruction template), so it follows its own existing shape rather than the templates'.

Testing

1. bats tests/test_doctor.bats — 35 ok, 0 not ok (unchanged from feat(doctor): default to the whole installation, filter with --project/--type/--team #654; doctor.sh itself is untouched by this PR).
2. bats tests/test_dispatch.bats — 8 ok, 0 not ok (6 pre-existing + 2 new for the doctor case).
3. bats tests/test_delivery.bats — adjacent suite, all green.
4. Packaging smoke: ran install.sh into an isolated HOME, confirmed the deployed claude-code/template.md mentions doctor and the deployed scripts/windows/dispatch.sh routes -- doctor to the real doctor.sh, reporting a clean empty installation end-to-end.

[fujibee]

Why this cannot simply be retargeted to integration/remote

Releases are currently cut from integration/remote (rc.1 → rc.3), and this PR
plus #661 target main. Retargeting them as-is would document an interface the
shipping branch does not have.

main holds exactly one commit that integration/remote does not:

$ git merge-base --is-ancestor origin/main origin/integration/remote ; echo $?
1
$ git rev-list --count origin/integration/remote..origin/main
1
$ git log origin/integration/remote..origin/main --format='%h %s'
4d3202d feat(doctor): default to the whole installation, narrow with --project/--type/--team (#654)

That commit is the one that changed doctor's interface. The two branches parse
different arguments today:

$ git show "origin/integration/remote:scripts/doctor.sh" | grep -nE '^# Usage|--project|--redacted\)'
6:# Usage: doctor.sh <project_path> <agent_type> [--redacted]
52:    --redacted) REDACTED=1; shift ;;
53:    *) echo "doctor: unknown option: $1" >&2; exit 2 ;;

$ git show "origin/main:scripts/doctor.sh" | grep -nE '^# Usage|--project\)'
6:# Usage: doctor.sh [--project <path>] [--type <type>] [--team <team>] [--redacted]
65:    --project)

So on integration/remote, --project / --type / --team are rejected as
unknown options, and the positional project path and agent type are required.

This PR's wiring — and #661's README/llms.txt text — describe bare doctor as
the normal form and the three flags as the way to narrow it. That description is
accurate for main and wrong for integration/remote.

Ordering

#654 has to reach integration/remote before the wiring or the human-facing
docs land there, or the documentation ships ahead of the implementation. Merges
in that direction already exist (#657, 0672a2c, 1f24501), so the mechanism
is routine; only the ordering matters.

Either sequence works:

1. #654 merges into integration/remote first, then fix(install): refuse to guess between multiple installs on --update with no --cmd #659/feat(doctor): wire into the 9 per-type templates, the Windows dispatcher, and the plugin-path SKILL.md #660/docs: cover doctor on the human-facing surfaces (README, README.ja, llms.txt) #661 retarget; or
2. fix(install): refuse to guess between multiple installs on --update with no --cmd #659/feat(doctor): wire into the 9 per-type templates, the Windows dispatcher, and the plugin-path SKILL.md #660/docs: cover doctor on the human-facing surfaces (README, README.ja, llms.txt) #661 merge into main, and the next main → integration/remote
   merge carries all of it together.

Both routes pass through a merge into or out of main, which is not something a
seat can perform.

Measured, not run

The above is read from the two trees' argument parsing. Neither version was
executed for this note.