fix(skills): report the engines with no skills primitive instead of dropping them

[clawedassistant26]

The bug

prd/0003 R8 says engines without a skills primitive MUST be reported as skipped, and the PRD's own UX Notes print exactly that:

moshcode skill install https://github.com/acme/some-skill
#   ✓ claude    cloned into ~/.claude/skills/some-skill
#   ✓ gemini    installed (--scope user)
#   · codex/opencode/aider  skipped — no skills primitive

planSkillInstall maps SKILL_ENGINES, so codex, opencode, privacycode and aider never enter the plan, never reach runSkillInstall, and never reach summarize. The user is not told they were skipped — they are told nothing at all.

A consequence worth stating: the default: return { skip: "no skills primitive" } branch in skillInstallAction is unreachable through the command. It is only ever called with keys from SKILL_ENGINES.

Reproduced through the real CLI

No hand-editing, HOME pointed at a scratch dir so nothing touched a real skills directory:

$ moshcode skill install https://github.com/acme/some-skill
· installing skill some-skill → https://github.com/acme/some-skill across skills engines…
   claude    not installed — /install claude
   gemini    not installed — /install gemini

Two rows. Four engines silently absent.

The asymmetry that makes it a bug and not a choice

/skill list already gets this right, and its comment states the principle:

The engines with no skills primitive are whatever ENGINES has left over. Derived, not hardcoded: a hardcoded list silently drops any engine added later, so the matrix stops showing every engine moshcode supports.

Same command family, same run, same machine:

surface	engines reported
/skill list (the matrix)	6 — claude, gemini, opencode, privacycode, codex, aider
/skill install (the fan-out) — before	2 — claude, gemini
/skill install (the fan-out) — after	6

privacycode is the concrete case the matrix comment warns about: an engine added after SKILL_ENGINES was written, present in the matrix, missing from the fan-out.

The fix

10 insertions, 2 deletions, one file. Plan over every engine, SKILL_ENGINES first then the remainder, derived from ENGINES the same way the matrix derives it.

runSkillInstall already checks skip before installed, so a skipped engine reports its reason whether or not the binary is present — matching runMcpAdd's existing precedent for R6.

After:

$ moshcode skill install https://github.com/acme/some-skill
· installing skill some-skill → https://github.com/acme/some-skill across skills engines…
   claude    not installed — /install claude
   gemini    not installed — /install gemini
   opencode  skipped — no skills primitive
   privacycode skipped — no skills primitive
   codex     skipped — no skills primitive
   aider     skipped — no skills primitive

Tests

New test/skill-install-fanout.test.mjs, 16 tests.

Seven are the bug: plan covers every ENGINES key; each non-primitive engine carries the skip reason; the fan-out reports them as skipped; skipped beats not-installed when nothing is installed; an installed non-primitive engine is skipped and never spawned (asserted by counting spawns); privacycode specifically; fan-out and matrix name the same set.

Nine are controls in the opposite direction, so the fix cannot buy green by over-reporting: SKILL_ENGINES still exactly ["claude","gemini"]; the primitive engines still come first in order; claude's git clone argv byte-identical; gemini's argv byte-identical; installed/not-installed still decides a real target; a signal still reports failed; a non-zero exit still reports failed and not a skip; no engine planned twice; every entry carries its real bin.

Fail-before / pass-after: 7 fail, 9 pass unpatched → 16/16 patched.
Suite: 622 → 638 tests, 0 failing (pass 499 → 515, skipped 123 unchanged).

A test that passes by accident

test/skills.test.mjs:49 — "skillInstallAction: engines without a skills primitive are skipped" — guards precisely the right function and passes unchanged, because it calls skillInstallAction("codex", …) directly, with keys planSkillInstall never passes it. The skip branch is proven correct and simultaneously unreachable in production. The new tests go through planSkillInstall/runSkillInstall, which is where the defect lives.

Deliberately out of scope

planMcpAdd maps MCP_ENGINES, which omits aider, so mcp install prints five rows where the PRD example shows a sixth (· aider skipped — no MCP support). Observed in the same session and not claimed or changed here: that is R6/R7 territory with its own per-engine reasons, and it deserves its own change rather than riding along on this one.