feat: give a finished session a record, and show it in Review

[PurpleSentinel]

Part of #137 and #142. Stage A of the session record: the record exists and is shown. Stage B
puts it on the card.

The gap

The device ran a full session and forgot all of it the moment you double-tapped.

Most of the fix was connecting parts that already existed. SessionReviewController takes
a SessionSummaryProvider; SessionSummaryV1 already carried duration, overrun, completion
reason and integrity. Nothing implemented the provider, so the router passed nullptr.

MemorySummaryStore is that provider — last eight sessions, newest first, refusing anything
that fails the summary validator so a malformed record cannot reach Review and be rendered as
though it were real.

Three gaps, all wiring

• imu_meter_.update(input, false) — that argument is session_active, hardcoded. Correct
  when written, because no session existed on the device; stale for three merges. Peaks
  described everything since boot.
• The router dropped vertical_g, computed by the calibration since fix(imu): track attitude with the gyroscope so tilt is not reported as G #146.
• SessionSummaryV1 had no G fields, so a scoped peak had nowhere to live.

Then Review turned out to be invisible

Reported as "no data in review". The probe showed the record was stored, the model populated
and the routing correct:

REVIEWPROBE after capture   status=1 count=1 title='SESSION 1 OF 1' duration='DURATION 00:18'
REVIEWPROBE entering review status=1 count=1 title='SESSION 1 OF 1'

Two separate faults:

1. Every data row used Typography::caption — 0.82 mm on this 311 PPI panel, against 10 mm
   for the countdown. The data was on screen the whole time at a twelfth the height. This is
   Driver-facing typography is too small to read on the physical panel #127, and it was not merely ugly: it was hiding a working feature.
2. The peaks never reached the view model. I added them to the record and the store and did
   not carry them to the screen, so the G data genuinely was not there at any size. My
   omission, and nothing asserted that the model carries what the summary holds.

Rows are 1.65 mm now with a 2.82 mm peak headline, and the peaks block occupies the space lap
rows will take once there is a receiver, so only one is ever up.

Decisions

• Up and down peaks are separate — a kerb throws the car up, a compression loads it down.
• Peaks reset when a session starts, not when a trigger arms. A driver waiting on a launch
  has not started; carrying the device to the car is not the session.
• Sessions are marked degraded_gnss — true, and Review should say so rather than present
  a lap-less session as complete.
• Old peaks are not migrated. Anything recorded before fix(imu): track attitude with the gyroscope so tilt is not reported as G #146 latched tilt as acceleration
  with left and right swapped. kImuMeterSchemaVersion → 2 and kLogFormatVersion → 3 refuse
  them by version: they are wrong, not old.

Verification

37 host suites (one new) and 70 simulator tests. New coverage for newest-first ordering,
bounded history, refusal of records that fail their own validator, peaks on every axis, and
session scoping including that peaks survive the end of a session — which is when they are read.

Confirmed on the panel across several sessions, ending with the record read back at 330 s
uptime and Review rendering it. Probe removed and its absence verified on the final image.

Deliberately not here

It is in RAM. Eight sessions, gone on power cycle. The card is Stage B, and that is where
the SD write path finally gets exercised before GNSS lands on top of it at 25 Hz.

#162, raised from this work: the recorded peak inherits the radar's 2 g display clamp. Safe
for the horizontal axes on a real car, not safe vertically — a kerb passes 2 g easily and would
be recorded as exactly 2.00, a clamp presented as a measurement.