Captain-answer attribution has no per-key record, so the gated-hold shape cannot be marked

[HelloWorldSungin]

Split out of #235 on 2026-09-04 as a deliberate deferral with a record, not a quiet scope reduction.

What #235's fix does and does not cover

#235's accepted fix makes a self-decided ask-user finding visible by marking firstmate's own answers, so an answered call can be told from one a worker decided for itself. That marking works only where the answer's decision key is recoverable from what already exists: the key IS the held task id, or it is the legacy <origin>-decision-<key> form that encodes the key in the id.

It does NOT cover the shape captain-hold-lifecycle itself recommends. That skill says, verbatim, "Prefer holding the work item the question gates over minting a new row", so the common case is an origin task escalating needs-decision [key=route] while the held row is the gated work item under a different id. In that shape the mark cannot be attributed.

Why it was not simply fixed

Established from the code during the #235 work rather than assumed: command_complete receives a flat list of hold ids with no keys attached, and pairs them with fold-derived keys only by writing the whole inventory on every key's line, so the transfer note records no key-to-hold mapping at all. Any hold in the inventory therefore matches every key that the same complete transferred. Nothing else associates a hold with a key.

An attempted fix inside #235 made this worse rather than better: it over-marked, writing a false answered record for one call using another call's ruling and then suppressing the real answer. That is a false record, which is strictly worse than a gap.

So closing this properly needs durable per-key attribution that does not exist today. That is new state, it was judged an expansion beyond #235's accepted intent, and it was declined there deliberately.

What #235 ships instead

Marking restricted to the provably attributable shapes, and for every other shape the row is VISIBLY SILENT: it reports that this shape cannot be attributed and why, rather than emitting either a verdict or an accusation. A check that names what it cannot see is preferred over one that quietly gets it wrong.

The question this issue holds

Should firstmate carry durable per-key attribution so a captain answer can be tied to the specific decision key it answered, covering the gated-hold shape?

Worth weighing when it is picked up:

• The benefit is real: the uncovered shape is the one the lifecycle skill actively recommends, so the blind spot sits over the common case rather than an edge.
• The cost is real too: this mechanism produced defects in five consecutive review rounds, and the one attempt to add correlation without new state produced a false-record defect. New durable state on that path deserves a deliberate decision rather than an accretion.
• A cheaper alternative may exist and was not explored: having complete record the mapping at write time IF the caller can supply it, rather than reconstructing it afterwards. That was ruled out for A steer delegated ask-user finding authority back to the implementation worker, and nothing in the fleet noticed for four findings #235 only because the current caller does not have the keys, which is a caller question rather than a storage question.

Not urgent

Nothing is blocked. The shipped behaviour is honest about its own limits, which is what #235 asked for.