doctor reports lost decisions for pending captures that never held one

[MongLong0214]

Found by running doctor on this repository.

warn  pending captures — 4 capture(s) can no longer apply — their base commit is no
      longer HEAD; oldest from 2026-08-13T12:50:26.870Z. A staged record binds to the
      tree it was prepared for and is skipped once that tree moves, so these decisions
      were never written to the history (#458)

The four it is describing:

NONCE     PHASE     RECORDS  VALIDATION  AGE   BASE      FLAGS
ba54a69c  prepared  0        -           3d    ff333338  stale
bd065cf2  prepared  0        -           3d    9c4d25aa  stale
464b4c13  prepared  0        -           3d    9c4d25aa  stale
e8eb71c6  prepared  0        -           3d    9c4d25aa  stale

Zero records, none staged. Nothing was drafted, so no decision was lost. The warning says one was — four times.

Where it comes from

src/commands/doctor/checks/capture-pending-backlog.ts already draws the distinction that matters:

const lost = stranded.filter((transaction) => transaction.phase === 'staged');   // :131

const detail = lost.length > 0
  ? `${lost.length} staged capture(s) expired before reaching a commit and were dropped` + …
  : `${stranded.length} capture(s) can no longer apply — their base commit is no longer HEAD`;   // :139-145

Both branches are accurate. The sentence appended after them is not conditional:

`${detail}; oldest from ${oldest}. ` +
  'A staged record binds to the tree it was prepared for and is skipped once that tree moves, ' +
  'so these decisions were never written to the history (#458)'   // :152-154

That explanation is written for the lost > 0 branch and is emitted for both. When lost === 0 it describes a staging that did not happen and decisions that did not exist.

Why it is worth fixing rather than tolerating

The comment at :102-105 records that #584 fixed the neighbouring version of this — the alarm counted consumed transactions and so was "loudest on the healthy path it exists to distinguish from". This is the same residue one layer along: the count is now right and the sentence explaining it is not.

The practical cost is that a repository with only abandoned drafts is told it silently lost decisions, and commitlore pending ls — the offered fix — shows RECORDS 0 and no way to recover anything, because there is nothing to recover. A reader either distrusts the diagnostic or goes looking for records that never existed.

Suggested shape

Move the explanation inside the lost > 0 branch, and give the other branch its own — an abandoned draft that never staged is a non-event, and saying so plainly is the accurate report. severity may also be worth revisiting for the lost === 0 case, though that is a separate judgement from the wording.

Verified against source at 3843ed0, not the bundle.