name the case, the reason, and the difference in test failures

[kacy]

three defects in std.testing made a table of cases unpleasant enough to be worth avoiding — which is the actual reason exhaustive negative coverage doesn't get written, more than any missing helper.

a failure didn't say which case failed. check() took a name but compared only ints; check_str only strings. anything else fell back to assert_eq, whose report is literally named "assert_eq" — over a loop of seven vectors it named neither the row nor the parameters, leaving two hex strings to diff by eye. check is now generic and carries the label, and check_str routes through it, so the two hand-specialized copies became one implementation.

a text mismatch reported the two lengths. got 412 chars, want 409 tells you nothing about a difference four hundred characters in. it now reports the first differing offset with a window of each side:

first differ at char 3: got "def", want "Xef"

a rejection couldn't say why. every negative test in the crypto modules is assert(r.is_err), which passes just as happily when the call fails for a reason nobody intended — a malformed fixture, a renamed field, a check that ran before the one under test. assert_err_contains(message, part) takes the error's message and the text it should mention. that one is a correctness gap rather than ergonomics: several existing negative tests would pass today if the failure moved to an unrelated cause.

the boundary this uncovered, now documented

while adopting the labelled check inside a colocated test block, the block's count silently dropped — std.testing keeps its own tally, so calling these helpers inside a test block records into a count that block never prints, and the assertion quietly stops counting. i caught it because a suite went from 10 tests to 8.

the module header now states it: built-in assertions inside test blocks, these in a test script that ends with done().

that also bounds what this PR buys. the labelled form helps tests/cases-style scripts today; colocated module tests still can't share a table-driven runner, because assert_eq is undefined outside a test block. that remains compiler work and is the next thing worth doing — but it is genuinely separate, and worth knowing that inline tables inside a test block work fine at one-module scale (the argon2 and jwt vector tables in #642 are written that way).

what was tested

• tests/cases/test_std_testing_helpers.pith exercises each new assertion, including the labelled form reporting by case name and the text diff; expected output refreshed
• make run-regressions-only: 304 passed, 0 failed
• make run-examples: 105 passed, 0 failed — examples/stdlib_test.pith output byte-identical, confirming the generic check is compatible with the 311 existing int call sites
• tests/cases/test_suite.pith (the heaviest check consumer): 46 passed
• doc coverage clean

notes

• check_* names are kept rather than deleted. removing them would mean rewriting 400+ call sites across four files for no user-visible gain; unifying the implementation behind them gets the dedup without the churn.