fix: first-run FDA assist survives a jq-less machine (#394)

[kiki830621]

Refs #394

What

1. plugin/hooks/session-start.sh — jq/ps dependency gates 從檔案頂端下移到 first_run_fda_assist 呼叫之後、staleness 區塊之前。FDA assist（mail#355）不需要這兩個工具，卻被擋在 gate 後面 —— 沒裝 jq 的機器（恰是 fresh-install 的目標受眾）整個 assist 靜默消失。staleness 區塊的 graceful-skip 語意逐 byte 不變。
2. plugin/tests/test-session-start-hook.sh — 兩個測試缺陷修正 + FDA 分支補洞（22 → 42 asserts）：
   • Case 5 隔離修正：舊 PATH（/sbin:/usr/sbin）同時移除 jq 和 ps，「jq missing」case 在只有 ps gate 正常時也會過。改用 symlink-farm shim（make_shim_without jq）真正只抽掉一個工具。
   • FDA 分支從零覆蓋到五個 case（Case 9–12 + 核心回歸鎖 Case 13/14）：assist 觸發（marker + detached --setup 探針）、once-only、FDA 已授權不擾、舊 binary 版本閘跳過、jq 缺席時 assist 仍要跑（bug: session-start hook 的 jq/ps gate 擋在 first-run FDA assist 之前 — 沒 jq 的機器 FDA 協助永遠 no-op (#335 verify) #394 的存在理由）、ps 缺席對偶 case。舊 suite 的 22/22 之所以蓋不到這些：HOME override 讓 [ -x $HOME/bin/CheAppleMailMCP ] 恆假、assist 每個 case 都自我跳過。

Verification

• bash plugin/tests/test-session-start-hook.sh — 42 pass / 0 fail（既有 8 case 全數保持綠 = staleness 行為回歸鎖）
• bash -n hook + test 皆過

過程中抓到的實作細節

gate 下移讓 PLUGIN_ROOT 的 dirname 先於 gate 執行 → shim 工具清單必須含 dirname（第一輪 shim 漏掉，Case 5 的 stderr 潔淨斷言當場抓到 —— 隔離測試自己證明了自己的價值）。

---

🤖 /idd-all batch lane 4/6。Do NOT add a GitHub close trailer — issue lifecycle 走手動 close skill。

[kiki830621]

6-AI verify — round 2 + fixes (#394)

Backend: pai-ensemble 2.20.0 · 4 lenses + devil's-advocate + Codex gpt-5.6-sol (xhigh) · dispatch model opus · 46 findings (2 HIGH / 16 MEDIUM / 18 LOW / 10 INFO).

The two HIGHs were both demonstrated by mutation, not argued — and both showed round 1's regression locks were assertion theater.

The two that mattered

1. The gate-existence lock pinned the diagnostic string, not the gate.
Replacing both gate exit 0 with : — keeping command -v and the debug echo — left the suite 46 pass / 0 fail. The message still printed; execution simply fell through and failed harmlessly a few lines later.

2. Cases 13/14 never asserted that --setup actually launched.
wait_for_file … || true discarded its own result. Making the setup launch jq-dependent — #394's exact regression — also left the suite fully green, because the cases asserted the assist's three stderr lines and never its actual product.

The second is the sharper one: #394's value proposition is "a jq-less machine still gets the FDA assist", and the assist is the setup window, not the text.

What changed

• The staleness block now lives in run_staleness_detection() with return-scoped gates. The scoping is structural, not positional — a feature added below it still runs without jq/ps. Ordering was bug: session-start hook 的 jq/ps gate 擋在 first-run FDA assist 之前 — 沒 jq 的機器 FDA 協助永遠 no-op (#335 verify) #394's root cause; leaving bare exit 0 left the next feature to re-inherit it, enforced only by a comment.
• A debug-gated staleness gates passed seam below the gates locks their effect. Cases 13/14 assert it is absent; new Case 15 is its positive control, without which the absent-assertions would pass vacuously if the seam were deleted.
• Cases 13/14 now assert the launched setup window and the once-only marker.
• CHE_MAIL_HOOK_DEBUG is read as exactly "1". Under -n, the value 0 — which every reader takes as off — turned debug output ON, in real sessions too. run_hook neutralizes an inherited value; new Case 16 locks it. Before this, CHE_MAIL_HOOK_DEBUG=0 bash … made the suite 45/1.
• The probe mock validates full argv, so --quietly no longer satisfies the --check-fda --quiet contract the version gate exists to protect.
• Case 12 asserts an old binary is never probed at all — the harm the gate prevents — rather than merely that no message appeared.
• make_shim_without post-checks that exactly one tool is missing; its silent continue for builtins could otherwise thin the shim back toward the both-missing state Case 5 was fixed for.
• Unset HOME under set -u would abort with a non-zero exit, the one thing this hook promises never to do.

Evidence

bash plugin/tests/test-session-start-hook.sh          → 60 pass / 0 fail  (16 cases; was 22)
CHE_MAIL_HOOK_DEBUG=0 bash …                          → 60 pass / 0 fail  (was 45/1)
bash -n  (hook + suite)                               → clean

Five mutations run, five caught — including the two that previously survived (gates that print but do not return; a jq-dependent --setup launch).

Filed, not silently folded in

--check-fda --quiet is a four-value contract (SetupCLI.ExitStatus: 0 granted / 1 denied / 2 noMailData / 3 undetermined) that the hook collapses to a boolean. noMailData — the normal state of a machine where Mail has never been opened, i.e. exactly this assist's audience — makes it report "Full Disk Access is not granted" (unfounded) and burn the once-only marker, so the assist self-disables precisely where it is most needed. Pre-existing since mail#355 and outside #394's gate-ordering scope → #403, with a pointer in the hook.

Honest status

Not independently re-verified. MAX_ROUND=2 was reached; the mutation battery is self-verification, not an adversarial round.