v0.37.1

Hopping between related arias no longer re-reads what the screen already
holds. The daemon answers figaro.lineage (chain, ancestor, divergence
turn), the client keeps every turn below the divergence and reads only the
rest, and a hop back onto a parked aria costs a few hundred bytes with the
screen byte-identical. Measured on one honest corpus: parent to branch
84,720 -> 11,190 bytes, branch to parent 133,085 -> 4,021, a hop back while
parked on the divergence 136,871 -> 536; time to first paint on a forward hop
a tenth of what it was. plans/prefix-retention.md and
plans/transcript-benchmarks.md carry the numbers and the eight defects the
instruments found on the way, including a read loop that spent the whole
saving back while every per-hop number looked excellent.

Form deltas draw as adornments of the node they landed on: one Δ in a new
right gutter, d opens one row per key, t the tool body, Enter both, each
row selectable, joined by a snake cursor; a fork lifts into the inquiry's
header as ⑂ parent.

The restart-with-tool-in-flight corruption is fixed on both sides: a call
whose arguments never parsed leaves the IR and the native cache together,
and a pairing gate repairs histories already on disk, now without dropping
the real results of parallel tool calls. The four P1 findings of the earlier
review are closed (terminal controls in form values, nondeterministic delta
folding, clamped tool-output offsets, a request-global dressing), and the
quote walk copies only the two windows it sends (1,600 bytes at 4 MiB).
Pit rows carry their aria id as data, the pager no longer freezes on a
long open message, :fork follows the shell's rule for attendance, and the
bundled skill teaches all of it in reference/quoting.md.

The marks sink (FIGARO_MARKS) records keys, hops, reads, frames and memory
so the transcript can be measured like a web page; FIGARO_CLI_PPROF profiles
a session.