Record what the Live window actually withholds (#73)

[jerome-queck]

Closes #73

ADR-0002's list of Gated streams was Formula 1's 2025 announcement, taken on trust because nothing
else existed, and the record ended by saying so — one measurement outstanding. Zandvoort 2026
Practice 1 took it: captured live on an unauthenticated connection and then backfilled, the same
Session with the Live window's gating and without it. Three of the four named streams are wrong.

The decision is untouched. This project still holds no credential, and the argument for that is
the one it always was. What this rewrites is the account of what the decision costs — including the
part that got worse.

What changed, and why

• docs/measurements/what-survives-a-live-session-unauthenticated.md is filled in. Both tables
  are re-readable from the kept capture at .archive/live-captures/2026-08-21_Zandvoort_Practice_1/
  (outside the repository, so not committed): 22,233 SignalR messages over the hour with zero
  CarData.z and zero Position.z frames while subscribed to both, and 7,238 records live against
  819,984 in the Backfill of the same hour.

• docs/adr/0002 is revised in place, with a dated note at the top saying so and why — the
  decision holds, so this is not a supersede, and the note is what keeps the original reasoning
  legible. It now carries what is measured gated (telemetry, locations, both whole streams), what
  is presumed gated (standings, on the 2025 announcement's authority alone), what is measurably
  free (pit stop durations, 75 of 97 records live; the other 22 are lap-one pit exits with no stop
  to time), and the honest admission that the cost is larger than the record assumed and still the
  smaller of the two.

• The strip lists three Gated streams, not four — Positions, Telemetry, Standings. Stop times came off because they arrive. DRS came off because the 2026 regulations abolished DRS and
  car_data.drs is null in every 2026 reading, so the label named nothing. Nothing was invented
  to replace it: the capture shows OVERTAKE ENABLED / OVERTAKE DISABLED arriving as ordinary
  RaceControlMessages, on a stream that is not Gated and that the strip already draws two blocks
  up. A fourth chip would be a label for a stream that is neither missing nor a stream.

• The list is written in three places, so test/design-system.test.sh now holds both
  design-system specimens to the GATED constant the running strip renders, the way it already
  holds the two copies of the Driver-row header to each other. test/session-strip.test.sh names
  the three rather than only asserting the container exists.

• Smaller corrections that were left false by the above: CONTEXT.md's Gated stream entry,
  README.md, the drs comment in server/openf1/records.ts, and the "still unanswered" sentence
  in docs/measurements/a-race-session-on-disk.md — whose own 2025 table stands, because it
  measured a season that still had DRS.

• docs/adr/0012 gets the footnote the issue allowed it: upstream's starting_grid is empty for
  2026, so the scrape kept on the table as a fallback would currently return nothing to parse.

Left open

Standings are still untested, deliberately. Both championship collections are empty live and
empty in the Backfill, because a practice Session has no standings either way — that is a Session
type with nothing to say, not a measurement, and no conclusion is recorded from it. They stay on the
strip and in the ADR on Formula 1's word, and ADR-0002's Revisit when now names the race-day
capture that settles it.

The Session countdown is in the same position for a smaller reason: no ExtrapolatedClock frame is
in the capture, and the subscription list is not recoverable from the file, so "not sent" and "not
subscribed" cannot be told apart. The flag condition is answered — ten TrackStatus transitions
arrived inside SessionData.

Drift

#73 — ADR-0002 is out of date: the live window was measured and it withholds a different set: 0 of
0 criteria delivered, because a Decision issue carries no acceptance boxes. All five scope items are
done, and two things go beyond the brief.

➕ Beyond the brief — test/design-system.test.sh now holds both design-system specimens to the
GATED constant the running strip renders. The list is written in three files and has just moved
once; this repository already tests exactly this kind of drift for the Driver-row header, and the
alternative is a convention. Say the word and I'll take it out.
➕ Beyond the brief — README.md, the drs and car_data comments in
server/openf1/records.ts, and the closing sentence of
docs/measurements/a-race-session-on-disk.md are corrected too. The issue did not name them, but
each was left stating something false by the corrections it did name. Say the word and I'll take
them out.

Nothing was left undelivered. Standings staying untested is the issue's own instruction, not a
shortfall.

Checks

test/run, npm ci && npm run typecheck and shellcheck -x all green. No bin/ command was run:
this was written from a worktree, where the wrappers compute container paths from their own location
(ADR-0004) and would resolve wrongly.

Assisted-by: Claude Opus 5 (reasoning: high)
Co-authored-by: Claude noreply@anthropic.com