20260811 - Antenna geometry optional, cpi_s withdrawn, Q8 closed

[Purple10101]

Three commits, each a decision that reduced scope rather than added it.

---

6307e8a — Withdraw the cpi_s request

Q3 was about to ask the server author to add cpi_s to NodeConfig. Josh pushed
back — "cpi_s is 0.5s on the whole fleet, what does it buy?" — and neither
justification survived checking.

The one it leaned on hardest was gap detection: t[n+1] − t[n] > cpi_s reveals
dropped capture, which nothing else shows because seq stays contiguous through
it. Tested against real captured frames, it fires on 27 of 27 gaps — minimum
0.866 s against a 0.5 s CPI. Processing exceeds the CPI on every node, so it
describes the fleet's permanent state rather than an event.

The offset-correction argument was thin too: cpi is 0.5 fleet-wide, so it is a
constant you state once in prose, worth ±0.25 s inside a 4 s association window.

What Q3 keeps is the half that costs nothing — which edge is t? — plus the
fact the question surfaced: the node has no send cadence at all. One POST per
frame blah2 produces, no timer, no batching, at most one in flight. The arrival
rate at the server therefore is the radar's frame rate, so "2 Hz, fixed" cannot
be honoured structurally, not merely because processing is slow today.

2455304 — Close Q8

This service does not communicate tracks. Moved to Answered rather than deleted,
because the answer shapes the architecture: a second must-not-lose stream is a
different transport discipline from latest-wins and would have meant a spool.
Worth a later reader seeing it was chosen rather than overlooked.

5f14d1c — Antenna geometry optional, and wire/serialise.py retired

Both beam fields leave NodeConfig.required, agreed with the server author.
An uncharacterised antenna omits both keys and nothing is substituted — the
same discipline as the consent records. retina-gui is not collecting the geometry
for the foreseeable future, so absent is the fleet-wide steady state and the
docstrings say so rather than treating it as temporary.

Two earlier designs are recorded in Q1 so they are not re-derived: raising
IncompleteConfig (correct while the field was mandatory) and defaulting to 360
(still a claim the node had not made).

wire/serialise.py goes with it. It existed for exactly one field —
beam_azimuth_deg was required and nullable, so null was a value the server
expected rather than an absence. With no such field left, to_wire and
exclude_none=True became the same function.

The accepted risk, made loud

A future revision reintroducing a required-nullable field would break payloads
silently. tests/wire/test_payload_encoding.py guards it, and both canaries
were verified to fire against the regression they guard, not merely to pass:

• No payload field is required and nullable. Compares against the union's
  members rather than the annotation's text, because adsb_hex is
  list[AdsbHexItem | None] — nullable items inside a list, which
  exclude_none never touches. That distinction fooled a first attempt, so it
  has its own test.
• Every payload survives json.dumps. mode="json" is load-bearing and no
  longer owned by a helper: without it the acceptance timestamps stay as
  datetime objects and registration fails at send time.

One guard widened, not removed

_send_config caught IncompleteConfig to report a bad config instead of
killing the config loop. It is now except ValueError — and since pydantic's
ValidationError is a ValueError, it also catches a config violating the
spec's own bounds. A latitude past 90 or an fc_hz below the minimum previously
took the loop down.

---

Verification

Run against a tracked-files-only tree, because the working tree carries
gitignored drafts CI never sees — the check that would have caught two earlier
false passes on this branch:

• tests, ruff check/format, the dead-code gate, and the model drift check all pass
• all four beam-field combinations round-trip 200 through the mock on both
  register and config — neither field, width only, azimuth only, both

Note for the next spec revision

docs/node-ingest-v1.yml now carries a change that is ours, not theirs. If
the server author sends another revision it will not include it unless they
applied it upstream too — check NodeConfig.required when adopting one. The
exception is named in CLAUDE.md's working agreements and recorded in Q1.

🤖 Generated with Claude Code