fix: report receipt-less ok observations as unverifiable in receipts mode

[Abhishek-B-R]

Problem

The as-published day files currently carry receipts: [] on every observation, for example:

curl -s https://data.getcomputable.com/H100/published/observations/2026/09/23.json | jq '[.data.observations[] | (.receipts | length)] | unique'
[0]

The same holds for B200 on 2026-09-01, 2026-09-10 and 2026-09-23. The versioned day files (H100/v7/observations/2026/09/23.json) and latest.json still carry receipts.

./reproduce --receipts reads the as-published files by default, rebuilds zero passing sources and reports every printed value as a mismatch. Re-checked today on main (0530437):

$ ./reproduce --receipts h100 2026-09-23T12
H100 2026-09-23T12:45 recomputed NO-PRINT published 3.522611 (band 0.545911) MISMATCH digest OK version 7 methodology_id h100_sxm_v1_calc_v16
  published status ok but only 0 observed passing disclosed sources (< min_sources_to_publish 5): no composite is recomputable
summary: 4 observation(s): 0 MATCH, 4 MISMATCH, 0 degraded
(exit 1)

Expected: no MISMATCH. The published values are fine: ./reproduce --receipts --version 7 h100 2026-09-23T12 reads the versioned file and gives 4 MATCH, exit 0 (and --version 5 b200 2026-09-10T03 the same). An empty receipts array means nothing was disclosed to recompute from, which is a could-not-verify outcome rather than a disagreement.

Fix

• published/verify.py: an ok observation with an empty receipts array gets a new unverifiable verdict with the message "observation carries no receipts". No-print observations and every matching path are unchanged.
• scripts/verify_published_record.py: prints UNVERIFIABLE for it, appends , N unverifiable to the summary only when N > 0 (so the summary of every passing run, and the live test regex that parses it, stay byte-identical), and exits 2 like the other could-not-verify outcomes. Without --version it also suggests rerunning with the version shown on each line.
• README: one sentence in the --receipts paragraph.

A question for you: would you rather --receipts read the versioned receipts instead, the way #54 made --full fall back to them? That would verify these rows rather than report them as unverifiable. It is a larger change to the reader routing, so this PR keeps to the minimum, and I am happy to take it that way if you prefer. If omitting receipts from the as-published rows is not intended, the verdict change still seems right for any future receipt-less row.

Tests

• tests/unit/test_published_verify.py: the verdict for a receipt-less ok observation.
• tests/unit/test_published_reader_cli.py: exit code, summary and message. Both fail on main.
• pytest: 1711 passed, 22 deselected (live). ruff check src tests scripts: clean.
• Live, on this branch: ./reproduce --receipts h100 2026-09-23T12 reports 0 MATCH, 0 MISMATCH, 0 degraded, 4 unverifiable, exits 2, and prints the --version hint.

The commit is signed off per CONTRIBUTING.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}