fix(cli): make the refusal-resolution ratchet enforce the runnable-command grammar the benchmark already classifies #3595

📝 Bug Description

Two grammars in this repository define what counts as a "runnable continuation" for a refusal message, and they disagree today:

1. The static ratchet (internal/cli/refusal_resolution_ratchet_test.go, refusalRatchetNamedContinuationRegexp, ~line 151) matches only gentle-ai [a-z][a-z-]* — product name plus one verb, nothing after. A refusal printing gentle-ai review status --cwd <repo> passes this check.
2. The benchmark classifier (bench/classify.go, HasRunnableCommand with placeholderRun = <[^>\s]*>) requires at least one real argument and rejects unfilled <placeholder> tokens. The same message is classified not runnable there.

So a message the benchmark classifies out-of-band satisfies the ratchet that CI runs as a required merge check. The ratchet's own header states the limitation plainly: "it cannot catch a WRONG named continuation — one that parses, runs, and does not help."

This seam is the shared root of a recurring class of reports where printed guidance cannot be executed as printed:

• review: the defect-report scrubber redacts the product's own contract identifiers, making its printed commands unrunnable #3443 (scrubber destroys contract identifiers) — fix in flight via PR fix(review): preserve public contract identities in defect-report scrubber #3456
• review: the pi transport refusal never names the handshake that would fix it, and offers to disable review instead #3440 (pi transport refusal never names the handshake)
• bug(doctor): sync remedy is unrunnable for dangling config symlinks #3557 (doctor sync remedy unrunnable for dangling symlinks) — fix in flight via PR fix(doctor): avoid sync remedy for dangling config symlinks #3561
• fix(review): review status silently ignores --lineage, so the escalated abandon refusal names an unusable continuation #1997 (escalated abandon refusal names an unusable continuation) — related fix in flight via PR fix(review): scope status inventory by lineage #3268
• bug(sdd): the SDD task-failure continuation names the OpenSpec dispatcher unconditionally and omits the change selector #2790, fix(review): emit a runnable untracked-scope recovery transition #2895 (further continuations awaiting design/review)

And per #2834's measurement, internal/sddstatus/runtime_ledger.go alone declares 36 refusals with no runnable exit.

🔄 Steps to Reproduce

1. Observe refusalRatchetNamedContinuationRegexp = regexp.MustCompile(gentle-ai [a-z][a-z-]*) in internal/cli/refusal_resolution_ratchet_test.go.
2. Evaluate it against a refusal text such as run gentle-ai review status --cwd `` — it matches (gentle-ai review), so the site passes the ratchet.
3. Run the same text through bench/classify.go HasRunnableCommand: placeholderRun finds <repo> and the classifier returns false.
4. Result: the two enforcement layers disagree about the same emitted sentence, and CI's required check is the weaker of the two.

✅ Expected Behavior

The static ratchet enforces the same printed-command grammar the benchmark already enforces:

• a named continuation must carry real arguments, not stop at the verb;
• an unfilled <placeholder> tail is accepted only when the site explicitly declares operator knowledge via the existing refusal:by-design operator-knowledge: marker (or holds a baseline row);
• the baseline regenerates once at the switch and may only shrink afterwards;
• existing CI (already a required merge check) then enforces the stronger grammar going forward — no new CI lane needed.

Both command splitters needed already exist and are deliberately mirrored (internal/cli/review_printed_command.go SplitPrintedCommandWords and its bench twin); no third splitter should be written.

❌ Actual Behavior

The ratchet accepts verb-only matches regardless of what follows, including tails the benchmark classes as not runnable. Wrong or dead-end continuations therefore pass the required merge check until someone files a per-instance report.

🌍 Environment

• Current main (verified against source on 2026-08-22)
• Evidence gathered on Windows; the grammar seam itself is platform-independent

💡 Proposed Scope

Static unification slice only: swap the ratchet's continuation grammar to the shared printed-command grammar, add the operator-knowledge marker path, regenerate the baseline once. Estimated ~150–250 changed lines including tests; one revertible PR from current main. The dynamic half ("follow the printed recipe until the block clears") already has a canonical pattern (internal/cli/review_abandon_message_test.go) and grows case-by-case afterwards.

Non-goals: no execution sandbox in the unit lane, no new wire vocabulary, no parallel command splitter, no change to bench classification semantics.