feat(delegate): 清單加一欄記唯讀盤點的實際指令

切片交的那幾支，助理現在只會提醒不會動手，因為沒有東西記得下「那一段唯讀盤點到底要跑什麼」。清單加第十二欄記具體指令，種入那一支有值就拿它當動作、沒值才退回只提醒。

只認三種寫法。一行指令，路徑一律寫成代入點開頭，由種入那一支代進字面絕對根目錄；指令裡不可以出現金錢符號或波浪號，那兩種在無人值守那一輪解不出來也進不了允許清單，會被靜靜擋掉。要標未接線就寫理由，留白的話下一輪分不出是刻意還是漏填。沒有唯讀入口的寫減號。

觸發型與不交的列一律填減號。觸發的意思是呼叫整支技能，這一欄填了指令會讓種入那一支改拿指令當動作，於是整支交出降級成只跑一支腳本，該寫的頁一頁都不會寫，而且看起來完全正常。這條由檢核腳本擋。

逐項回各存放庫核對，不照抄既有盤點的措辭，因而抓到五處對不上實際腳本的地方。

最要緊的一處：既有盤點點名的那支工具腳本要三個參數，其中兩個無人值守那一輪根本拿不到，而且它是唯一會去問遠端的一方。改用同一件事的本機那一半，而那個子命令剛好在閘門腳本「免讀標準輸入」的清單裡。這一點非確認不可：同一支腳本的另一個子命令不在那份清單上，標準輸入是管線又沒人關閉時會一直等，實測會無限卡住。腳本自己的註解就記著這個坑，說工具腳本轉呼叫時曾經整支卡死——那正是先前心跳斷掉的同一種死法。

另一處：既有盤點點名的同步子命令會寫檔，不是唯讀。那一列剛好是觸發型所以填減號、衝突沒落地，但措辭本身是錯的。

七項標成未接線，理由都是要連網要金鑰。連網那一種一過期就讓那一項每輪失敗，或每輪靜靜回報沒事——後者更難查；而純本機讀取本來一輪都不會失敗。先填會連網的，等於用一批每輪報錯的項目把真的發現蓋掉。

還記下一件接線那天要注意的事：查遠端版本那一支永遠回成功，查不到就安靜放行，所以金鑰失效時它會靜靜回報沒有新版。接線要用另一個子命令，那個有「查不出來」這第三種結論。

檢核腳本從四項檢查改成五項。填錯欄位、指到不存在的腳本、用了認不得的代入點都算缺失；標未接線只印提示，那是判過知道還沒接，不是漏填；指到的 domain 本機沒裝也只提示，那是機器少裝一套不是清單填錯，算缺失會讓半套機器每輪亮紅。

五支技能異動流程一併改：寫清單時要問到並寫下這一欄，重判時要一併重問，刪除時要檢查別列有沒有指到一起刪掉的腳本，一次改多支時要注意跨存放庫搬腳本是這一欄最容易過期的地方。

skills/skill-check/SKILL.md

@@ -23,9 +23,9 @@ Single source of guidelines: [`../../references/guidelines.md`](../../references
    6. Run the two wiki-rule checkers once each, not per domain — both judge shared rules, so a second run adds nothing:
       - `jsc-gitea/tools/check-wiki-rules.sh`, which verifies wiki repo resolution and the `hash-id` rule for every page type. It takes no argument. Route each exit code: 0 — printed `OK` on stdout, every item passed; 1 — the first mismatch is printed on stderr as `{項目}: want=… got=…` and the script stops there, so report that item and rerun after the fix, because the remaining items were never reached. Those are its only two codes. Until this audit, no flow in the whole repository ever called it.
       - `tools/check-page-name.sh {root}`, where `{root}` is the directory holding the domain repos — the parent directory of the paths `tools/sync-domains.sh` printed in step 1, so no extra derivation is needed. It compares the page-name pattern in its three copies: `jsc-gitea/tools/page-name.sh` (the canonical one), `jsc-hooks/hooks/comment-scope.sh` and `jsc-log/tools/worklog-pending.sh`. Route each exit code: 0 — the three agree; 1 — the mismatches are printed on stderr as `{檔案}:{說明}`, so report each one as a compliance failure, and a copy that could not be found is one of those lines; 2 — usage error, the tool takes exactly one argument; 3 — none of the three copies was found, so the root is wrong: fix it and rerun. Record exit 3 as 「什麼都沒查」; it is **never** a pass. The three copies stay separate on purpose — a hook must be self-contained and may not depend on another plugin's path at run time — so consistency is checked here instead of shared in a function.
-   7. Run `tools/check-delegate.sh {root}` once for the whole round, with the same `{root}` item 6 passed to `check-page-name.sh`. It compares the delegation list `tools/delegate-spec.tsv` against the skills `tools/list-skills.sh` finds on this machine: one skill one row, eleven columns, every mandatory column filled, and every `next` naming a skill that exists. It belongs in group 1 for the same reason `ste100-lint.sh` does — it is a deterministic script verdict, and it is judged **once for the whole round** rather than per domain, because the list is a single file covering every domain. Handing it to the group 2 sub agents would have ten agents run the same script over the same file and report ten copies of the same lines, with no single verdict anywhere; handing it to group 3 would turn a pass-or-fail check into a suggestion. Route each exit code:
-      - 0 — the list and the machine's skills correspond one to one and every mandatory column is filled. **A run that printed lines on stdout and exited 0 passed.** Those lines are hints, not compliance failures, and they are printed on stdout precisely so they are told apart from the failures on stderr: `origin=seed` marks a row seeded from the earlier inventory that has not been through the decision tree yet, and a version-behind line marks a row whose recorded `version` trails its domain's current one. The version number is per domain, so one skill's change marks every other skill of that domain — counting those as failures paints whole domains red on every release, and the hint stops being read at all. Report the hint count and the rows, and open no decision-tree item for them.
-      - 1 — a missing row, a duplicate row, a row for a skill this machine does not have, an empty column, a column value outside its vocabulary, or a `next` naming a skill that does not exist. Every one is printed on stderr as `{清單路徑}:{domain}/{技能名}:{說明}`. Report each as a compliance failure, named by the skill it belongs to. A missing row means the assistant is blind to that skill; an extra row means it will trigger a skill that cannot be called, and a failing trigger retries instead of pausing.
+   7. Run `tools/check-delegate.sh {root}` once for the whole round, with the same `{root}` item 6 passed to `check-page-name.sh`. It compares the delegation list `tools/delegate-spec.tsv` against the skills `tools/list-skills.sh` finds on this machine: one skill one row, twelve columns, every mandatory column filled, every `next` naming a skill that exists, and every `probe` either a runnable read-only command, a `pending:{reason}`, or a `-` on the rows that take one. It belongs in group 1 for the same reason `ste100-lint.sh` does — it is a deterministic script verdict, and it is judged **once for the whole round** rather than per domain, because the list is a single file covering every domain. Handing it to the group 2 sub agents would have ten agents run the same script over the same file and report ten copies of the same lines, with no single verdict anywhere; handing it to group 3 would turn a pass-or-fail check into a suggestion. Route each exit code:
+      - 0 — the list and the machine's skills correspond one to one and every mandatory column is filled. **A run that printed lines on stdout and exited 0 passed.** Those lines are hints, not compliance failures, and they are printed on stdout precisely so they are told apart from the failures on stderr: `origin=seed` marks a row seeded from the earlier inventory that has not been through the decision tree yet, a version-behind line marks a row whose recorded `version` trails its domain's current one, a `probe=pending:` line marks a delegated slice whose read-only entry point is not wired yet, and a line saying a `probe` domain is not installed here marks a script this machine cannot check. The version number is per domain, so one skill's change marks every other skill of that domain — counting those as failures paints whole domains red on every release, and the hint stops being read at all. Report the hint count and the rows, and open no decision-tree item for them.
+      - 1 — a missing row, a duplicate row, a row for a skill this machine does not have, an empty column, a column value outside its vocabulary, a `next` naming a skill that does not exist, or a `probe` in the wrong shape — a command on a row whose `way` holds `invoke`, a `-` on a row whose `way` holds only `patrol` or `remind`, a dollar sign or tilde, an unknown substitution point, or a script that does not exist. Every one is printed on stderr as `{清單路徑}:{domain}/{技能名}:{說明}`. Report each as a compliance failure, named by the skill it belongs to. A missing row means the assistant is blind to that skill; an extra row means it will trigger a skill that cannot be called, and a failing trigger retries instead of pausing. A wrong `probe` fails every unattended round in the same silent way, and the command-on-an-`invoke`-row case is worse than a failure: the assistant runs a bare script where the whole skill was supposed to run, and the round looks clean.
       - 2 — usage error: the script takes at most one argument. Fix the call and rerun; this is a defect in this skill, not a finding about the skill set.
       - 3 — nothing was checked, because `tools/delegate-spec.tsv` is missing, the root could not be derived, or `list-skills.sh` listed no skill. Record it as 「無委派清單可查」 with the cause from stderr and carry it into the step 3 merge; **exit 3 is never a pass**, because a check that read nothing reports neither a missing row nor an extra one.
 
@@ -85,7 +85,7 @@ Single source of guidelines: [`../../references/guidelines.md`](../../references
    - Optimization options: apply / defer / custom. Record the chosen option in the finding's 決議 field as `套用`, `延後` or `自訂`, and today's date in 決議日期. Any suggestion that weakens a protection must name the protection it removes and must not be applied unless the user explicitly accepts that tradeoff. Cost optimization may move, merge, cache, or narrow checks; it must not delete a compliance check only because it is expensive.
 
    Completion condition: every domain's checklist is complete after the merge, with the two whole-round verdicts carrying the same value in every domain, and every compliance failure and every optimization finding has a recorded decision — every optimization finding carrying both 決議 and 決議日期.
-4. Apply the confirmed fixes and accepted optimizations — the file-change part MUST run as a sub agent, one sub agent per affected domain repo, and those sub agents run **in parallel**: each repo's files are independent. A fix that changes a skill's behavior also updates that skill's `## {name}` section in the same repo's `references/behaviors.md`, in the same pass, so the fix and the behavior list land in one PR. A confirmed `check-delegate.sh` fix is written by the **main agent**, never by the per-repo sub agents: `tools/delegate-spec.tsv` is one file for the whole skill set, and parallel agents writing one file overwrite each other's rows. A missing row is filled by running the decision tree of [`../../references/delegate-criteria.md`](../../references/delegate-criteria.md) for that skill through `jsc-ask:ask` and writing the answer as a row with `origin` set to `judged`; an extra row is deleted; a dead `next` is repointed at a skill that exists. A fix that changed a skill's behavior in this same round also re-judges that skill and moves its row's `version`. Then run `tools/sync-skill-manifest.sh {domain-path}` directly (no sub agent needed) for each affected domain repo to refresh that domain README's 「Skills 目錄」 section and bump the version in all three manifests. Route each exit code: 0 — the README block and all three manifests are synced; 1 — the domain path, `skills/`, `README.md`, the `JSC-SKILLS` markers, a `SKILL.md`, a manifest, or a manifest `version` field is missing, so fix the named cause on stderr and rerun; 2 — usage error, the script takes exactly one argument; any other code — the script runs under `set -e`, so treat it as an environment fault and stop, never as a successful sync. Completion condition: every affected repo carries the changes, the matching `references/behaviors.md` update for every fix that changed a skill's behavior, the `tools/delegate-spec.tsv` rows for every accepted delegation fix, and the manifest bump.
+4. Apply the confirmed fixes and accepted optimizations — the file-change part MUST run as a sub agent, one sub agent per affected domain repo, and those sub agents run **in parallel**: each repo's files are independent. A fix that changes a skill's behavior also updates that skill's `## {name}` section in the same repo's `references/behaviors.md`, in the same pass, so the fix and the behavior list land in one PR. A confirmed `check-delegate.sh` fix is written by the **main agent**, never by the per-repo sub agents: `tools/delegate-spec.tsv` is one file for the whole skill set, and parallel agents writing one file overwrite each other's rows. A missing row is filled by running the decision tree of [`../../references/delegate-criteria.md`](../../references/delegate-criteria.md) for that skill through `jsc-ask:ask` and writing the answer as a row with `origin` set to `judged`; an extra row is deleted; a dead `next` is repointed at a skill that exists; a wrong `probe` is rewritten per that same file — verified against the owning repo, not guessed — and set to `pending:{reason}` when the slice has no read-only entry point on this machine. A fix that changed a skill's behavior in this same round also re-judges that skill and moves its row's `version`. Then run `tools/sync-skill-manifest.sh {domain-path}` directly (no sub agent needed) for each affected domain repo to refresh that domain README's 「Skills 目錄」 section and bump the version in all three manifests. Route each exit code: 0 — the README block and all three manifests are synced; 1 — the domain path, `skills/`, `README.md`, the `JSC-SKILLS` markers, a `SKILL.md`, a manifest, or a manifest `version` field is missing, so fix the named cause on stderr and rerun; 2 — usage error, the script takes exactly one argument; any other code — the script runs under `set -e`, so treat it as an environment fault and stop, never as a successful sync. Completion condition: every affected repo carries the changes, the matching `references/behaviors.md` update for every fix that changed a skill's behavior, the `tools/delegate-spec.tsv` rows for every accepted delegation fix, and the manifest bump.
 5. Sync the canonical marketplace — a **required** step, never optional. The canonical pair lives in `plugins/meta` and every domain repo carries a byte-identical copy, so a fix that leaves the copies apart makes some repos register a stale plugin set. Run `tools/sync-marketplace.sh {domain} {repo-url} {description}` once with an existing entry's own current values (rewriting the same entry is idempotent); the script rewrites both canonical files and copies them into every domain repo. Route each exit code:
    - Exit 3 — written, but some domain repo is not present locally. Run `tools/sync-domains.sh`, then rerun this step.
    - Exit 2 — usage error: the script takes exactly three arguments. Fix them and rerun.