Integrate inline composer dictation - #31

…ed parser fix

Point packages/agent-transcript-parser at 7aae690, the merge commit of parser
PR #31. Its tree is identical to the branch tip the suites below ran against.
That merge carries the review fixes: a protection lift that takes only the
rungs it needs (the recorded transcript keeps its six newest tool outputs), an
idempotent input trim so one loss is never counted twice, and a Claude
projector that re-encodes only what the API accepts and heals transcripts the
old projector had already poisoned.

Review of the host side found that the disclosure this fix adds could not be
read. The success toast used the two-second default and clamps to three lines,
and the new clauses came last, so in a narrow pane they were exactly what the
clamp removed. Lossy switch toasts, pane and bulk, now stay up for ten seconds,
and describeShrink leads with what the user must act on, the omitted
attachments and the trimmed newest turns, before the counts in ladder order.
Every clause singularises; the lift often clears exactly one item, which made
"1 tool outputs cleared" the common case rather than the rare one.

A switch through the real planner whose only loss is an old pasted image now
pins truncatedBeforeSwitch and the toast line end to end. It is the one state
where clearedAttachments alone carries the flag, and removing the term from
the sum fails that test and nothing else did before.

The design doc gains the Claude image re-encoding, the still-unproven live API
acceptance, the subset-based lift and a "never copy a foreign attachment part
verbatim" rule. The 2026-09-05 spec gets As-built item 11 because the parser
header cites it and its rung numbers no longer match the code.

Refs #998, #999

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>