Record maximum G on all axes in every session and surface it in Review

[PurpleSentinel]

Objective

Every session — Track Day and Race, not only G-Only — must record maximum G force on
all axes, and those maxima must be available in Review afterwards.

The design consequence

logger::SessionSummaryV1 has no G-force fields at all:

schema_version, record_size_bytes, session_id, session_duration_ms, session_overrun_ms,
completion_reason, integrity, degraded_subsystems, lap_count, best_lap_index,
best_lap_duration_ns, source_*, logger_*

Adding maxima therefore means a log format schema change, which is the part that
needs designing rather than just coding:

• Bump kLogFormatVersion and add a SessionSummaryV2
• Decide whether V1 summaries stay readable. The reader already distinguishes
  SummaryReadResult::unsupported_version, so the machinery exists; the decision is
  whether old sessions remain reviewable or are rejected.
• record_size_bytes is already carried per record, which should make a
  size-tolerant reader feasible — worth confirming before committing to the approach.
• Decide which maxima are stored. ui::ImuPeakSummary already tracks acceleration,
  braking, left, right, total, and the position at which total occurred. Storing that
  structure keeps one definition of "peak" across the live meter and the summary.
• Decide whether vertical G is included. board::ImuSample carries a Z axis but the
  meter is planar today, and kerb strikes make vertical peaks noisy and of
  questionable value.
• Decide whether the peak is per-session or also per-lap. Per-lap is more useful for
  Race and considerably more data.

Where the peaks come from

ui::ImuMeterController::update_peaks already maintains exactly these maxima and is
host-tested. The session recorder should consume that rather than computing maxima a
second time, or the live radar and the stored summary will eventually disagree.

This must run in Track Day and Race, not only G-Only. In Track Day the meter is not
displayed, but recording continues — the same principle as lap times, which are recorded
throughout and only withheld from the live display.

Scope

• Extend the session summary with G maxima behind a version bump
• Feed peaks from ImuMeterController into the session recorder in all modes
• Surface maxima in SessionReviewController and the review screen
• Ensure peaks reset per session and cannot leak across sessions
• Handle an absent or failed IMU by recording maxima as unavailable rather than zero,
  since zero is a legitimate reading and would be indistinguishable
• Update docs/DATA_LOGGING.md and docs/SESSION_WORKFLOW.md

Acceptance criteria

• A Track Day session records G maxima and they appear in Review
• A Race session does the same
• Maxima match what the live meter showed during the session
• An absent IMU yields "unavailable", not 0.00 G
• The version decision for existing summaries is implemented and documented
• Round-trip is covered by host tests

[PurpleSentinel]

The vertical axis now exists, and two upstream defects affect what this issue records

This issue asks for max G on all axes. Until now GravityCalibration::resolve returned
only lateral and longitudinal and discarded the vertical component entirely, so "all axes"
was not reachable from the data the device produced.

PlanarG now carries vertical_g, positive upward, alongside the horizontal pair. Kerb
strikes and compressions are a different kind of event from anything the tyres do in plan,
and they are the axis a driver cannot feel accurately, so they are worth recording
separately rather than folded into a total.

Two things to know before building the summaries, because both change the numbers:

Peaks were latching tilt as G (#143). The gravity reference was frozen at calibration,
so any later change of tilt read as acceleration at sin(angle) - 0.17 g on a 10 degree bank,
0.10 g sustained on a 10% gradient. Peaks are a running std::max, so those artefacts were
not averaged away, they were latched. Any maximum recorded before the gyro-tracked attitude
landed is suspect, and a steep pit-exit ramp could set a longitudinal maximum the car never
pulled.

Left and right were swapped (#144). The vehicle frame was built with "up" pointing down,
inverting the lateral axis, so ImuPeakSummary::left_g was holding right-hand load and vice
versa. Longitudinal was unaffected.

Consequences for this issue

• ImuMeterSnapshot/ImuPeakSummary need a vertical peak, which means bumping
  kImuMeterSchemaVersion from 1
• kLogFormatVersion still needs its bump, as already noted here
• Any peak data captured on the device before those two fixes should be treated as invalid
  rather than migrated

[PurpleSentinel]

The kLogFormatVersion bump this issue has been waiting on is done: #150 took it to 2, because session and rest durations widened from uint16 minutes to uint32 seconds and SessionSettingsV1 mirrors those fields. Logs written before it are rejected by version rather than silently misread as very short sessions. Adding the vertical G peak still needs kImuMeterSchemaVersion to move from 1, but no longer needs a log format change of its own.

[PurpleSentinel]

Plan, 2026-08-20: the record first, the card second

Starting work. Splitting it, because the two halves have different risks and the first
settles the shape the second has to persist.

Stage A - a session leaves a record. Peaks become session-scoped, the vertical axis is
plumbed through, a summary is captured when a session ends, and Review shows it. In memory
only, so it does not survive a power cycle.

Stage B - the record survives. That summary written to the card, read back, and several
sessions of history. This is where the SD write path finally gets exercised on device.

What is already there, and what is missing

The review pipeline is complete and waiting: SessionReviewController takes a
SessionSummaryProvider, and SessionSummaryV1 already carries duration, overrun, completion
reason, integrity and lap fields. Nothing implements the provider, so the router calls
review_controller_.begin(nullptr) and Review shows nothing.

Three concrete gaps, all in wiring rather than logic:

• imu_meter_.update(input, false) - the second argument is session_active, hardcoded
  false. It was correct when written, because no session existed on the device; three merges
  later a session does. Peaks are computed continuously and never scoped to anything, so
  there is nothing to put in a summary.
• The router builds the meter's input from resolved.lateral_g and resolved.longitudinal_g
  and drops vertical_g, which fix(imu): track attitude with the gyroscope so tilt is not reported as G #146 added precisely so kerbs and compressions could be
  recorded.
• SessionSummaryV1 has no G fields at all, so even a scoped peak has nowhere to live.

Schema consequences

• kImuMeterSchemaVersion 1 to 2, for the vertical current value and the up/down peaks
• kLogFormatVersion 2 to 3, for the peaks on the summary record

Peaks captured before #146 are not migrated. They were recorded against a frozen gravity
reference that latched tilt as acceleration, and with left and right swapped - they are wrong,
not merely old.

Scope for Stage A

• Vertical G through the meter: current value and separate up and down peaks, since a kerb
  strike and a compression are different events
• Session-scoped peaks: pass the real session state, so peaks reset when a session starts
• Peak fields on SessionSummaryV1, with the format bump
• Capture a summary at session end, from the session snapshot and the meter peaks
• An in-memory provider, so Review shows the session just finished
• Host tests for the scoping, the vertical peaks, and the captured summary

[PurpleSentinel]

Stage A merged in #163

A finished session now leaves a record and Review shows it: duration, overrun, and peak G on
every axis including vertical. The peaks are session-scoped, which they had never been - the
meter was still being told no session was ever running.

Two things worth carrying forward from building it.

Review was invisible rather than empty. The record was stored and the model populated all
along; every summary row was rendered at Typography::caption, which is 0.82 mm on this
panel against 10 mm for the countdown. #127 was not merely a styling complaint - it was
concealing a working feature. The rows are 1.65 mm now with a 2.82 mm peak headline, but every
other screen still ships the old scale.

Nothing asserted that the view model carries what the summary holds. The peaks reached the
record and the store and were never carried to the screen, and 37 suites stayed green. Worth a
test when Stage B lands.

Stage B, still open here

• Write the summary to the card, and read it back across a power cycle
• More than eight sessions, and history that outlives a reboot
• Exercise the write path under load: card pulled mid-session, card full, slow writes

That last point is the reason to do this before GNSS rather than after. The SD write path has
never run on this device, and a receiver will land on top of it at 25 Hz. Finding out then
whether it holds up is the worst possible time.

Also raised from this work: #162, the recorded peak inherits the radar's 2 g display clamp,
which is safe horizontally on a real car and not safe vertically.