Check the whole forced report, and skip visibly when it cannot be checked

[emrefbulut]

What changed

test_the_report_is_78_columns_of_ascii now checks every line of the --force output instead of stopping at the line ending in MEASUREMENT, and asserts the exit code and a minimum line count first.

The one assertion that genuinely cannot hold without torch — that the forced MEASUREMENT block is present — moves into its own test behind an explicit skipif.

Why

The old test looked like it covered the quotable report. It did not:

for line in text.splitlines():
    if line.strip().endswith("MEASUREMENT"):
        break                      # <- everything after this was never checked
    block_lines.append(line)

Its coverage silently depended on the environment:

• with torch — the forced measurement block exists, and the break skipped it
• without torch — the block does not exist at all, so there was nothing to skip

Either way the output --force uniquely produces had never been width- or ASCII-checked in any environment, and the test reported passed in both. Dumping a real run shows exactly what was escaping (lines 35–40):

32 [checked    ] started       no. This version of the command stops before training
33 [checked    ] ==============================================================================
34 [checked    ]
35 [AFTER-BREAK] FORCED MEASUREMENT
36 [AFTER-BREAK] stride=512  seed pairs=1  split=42  train=0
37 [AFTER-BREAK] inflation=+0.0 pp (uncertainty not estimated, n=1)
38 [AFTER-BREAK] recording-level: test 50.00%, window-level: test 50.00%
39 [AFTER-BREAK] environment: device cpu | torch 2.13.0+cpu | numpy 2.5.1 | scipy 1.18.0
40 [AFTER-BREAK]              sigmf 1.13.0 | cuda none

The exit-code and line-count assertions are there for the same class of reason: without them a run producing no output at all would satisfy every loop by iterating zero times.

How it was verified

The skip is now visible. Forcing _torch_installed() to return False — the state CI's test (3.11) / test (3.12) jobs are in:

SKIPPED [1] tests/test_preflight.py:323: the forced MEASUREMENT block only exists
once training can run; install the torch extra to exercise it
1 passed, 1 skipped, 17 deselected

skipped, with a reason, rather than a silently narrowed passed.

Mutations — and one of them corrected my own assumption:

mutation	result
unbreakable 95-character token in the measurement block	caught by both tests
non-ASCII character in the measurement block	caught by both tests
80-column breakable line in the measurement block	not caught

The third is the honest caveat and I am not going to paper over it. The measurement block is printed through rich, which wraps at the console width, so an over-long sentence is wrapped before it ever reaches the output and cannot violate the contract. The width assertion on that block is therefore load-bearing for content rich cannot wrap — paths, hashes, long numbers — and for ASCII. That is narrower than "this block is 78 columns", and claiming the wider thing would repeat the mistake this PR is fixing.

The preflight block above it is built with _field_lines and plain print, so there the width check is load-bearing in full.

Conventions it touches

• 2 A passing test does not prove it can fail — the whole PR. The old test had never been red on the block it appeared to cover, and my first mutation attempt showed my replacement was weaker than I assumed until I found one rich cannot wrap.
• 3 Do not claim what you did not measure — the caveat above, and the skip reason naming what is missing.
• 1 Never fall back silently — applied to tests: a check that cannot run must say so, not quietly check less.

Noticed, not fixed here

Line 32 above says started no. This version of the command stops before training and lines 35–40 are the measurement it then performed. That is the self-contradiction tracked as Bölüm C item 5; this PR only stops the test from ignoring that region.