Fix Obs80 reader emitting positionless orphan observations (deleted obs + S/s mis-pairing)

[matthewholman]

Summary

The obs80 reader let two kinds of malformed / non-observation lines reach the fitter as positionless records (sys='', pos=nan). For a space-based obscode (e.g. WISE / C51) such a row carries no ADES observer position and, downstream, falls back to a per-row JPL Horizons lookup — which 503-throttled under the high parallelism of the full MPC-catalog fit (surfaced on object j4767).

Two related bugs, both fixed here:

1. Deleted / replaced observations (MPC note 2 code X / x) were read as normal observations. j4767's WISE block contains one x line at the same timestamp as a kept satellite obs (a superseded measurement of the same exposure); tests/data/03666.txt contains a 1938 X photographic plate. These are now skipped.

2. Two-line records were paired purely by position. Satellite S / radar R / roving V first lines were merged with whatever line followed, with no check that the follower was the matching lower-case s / r / v continuation. A desynchronised file (a duplicated or orphaned continuation line, or a first line whose continuation is missing) could mis-pair or emit a positionless standalone.

Change

The three previously independent pairing loops (get_row_count, read_rows, read_objects / _build_id_map) are consolidated into a single _iter_records generator that is now the sole source of truth for how raw lines group into records. It:

• skips deleted observations (X / x),
• validates each continuation line against its first line (lower-case note 2 + matching designation cols 1–14 + matching obscode cols 78–80),
• drops orphan / mismatched continuations and continuation-less first lines instead of mis-pairing,

so every read path agrees on the record set (len(read_rows()) == get_row_count(), and read_objects stays index-aligned with _build_id_map).

Validation

Verified on the real j4767 bytes from the catalog run:

reader	rows	C51 rows	C51 positionless
before	452	13	1 (→ Horizons)
after	451	12	0 (all ICRF_KM)

Tests added:

• a real-data fixture (tests/data/j4767_wise_excerpt.txt, j4767's WISE/C51 block) asserting all 12 satellite records carry their ICRF_KM position and the deleted x line is dropped;
• synthetic deleted, duplicate-continuation, leading-orphan, missing-continuation, and count/read/objects-consistency cases (all four desync cases fail on the old reader, pass on the new).

All pre-existing Obs80Reader tests pass. tests/data/03666.txt's expected row count is updated 4313 → 4312 — the old value had counted the deleted 1938 plate as a real observation.

🤖 Generated with Claude Code

---

Update — subsumes #411

Folded in #411 (issue/402-407 Obs80Reader robustness), now closed as superseded. This PR keeps #411's convert_obs80 by-type two-line dispatch — S satellite ICRF, V roving-observer WGS84 geodetic (previously mis-read as a satellite position, #402/#282), R clear 'ingest via ADES' error — plus its _is_header_row/#407 blank-line guards and its roving/radar/blank-line tests + obs80_two_line_records.txt fixture. Combined with the deleted-X/x skip and S/s orphan fix here, _iter_records is now the single source of truth for record grouping across all read paths. Closes #402, #407. 18 Obs80Reader tests pass.

Closes #402
Closes #407

[kjnapier]

Looks good