Per-test cleanup reads the whole shared temp directory, so old leftovers slow every later run

[Chemaclass]

Problem

bashunit::cleanup_testcase_temp_files runs in every test's EXIT trap and expands:

matches=("$BASHUNIT_TEMP_DIR/${BASHUNIT_CURRENT_TEST_ID}"_*)

Expanding that glob makes bash read all of BASHUNIT_TEMP_DIR. That directory is shared (${TMPDIR:-/tmp}/bashunit/tmp) and persists between runs, and nothing ever reaps it — so every file an interrupted run left behind is re-examined by every test of every later run.

Worse, the scan can never match one of them: the id is built as <fn>_$$_<suffix>, so leftovers always carry a different PID. The work is pure overhead.

Measured

--no-parallel, one file of 100 trivial tests, varying only the number of unrelated files in BASHUNIT_TEMP_DIR:

leftovers	run time
0	498 ms
500	513 ms
2000	635 ms
5000	978 ms

Linear and unbounded — the framework gets slower the longer the machine has been used. Profiling a 200-test run against a real 4,242-file directory, this one call was 3.16 ms of an 8.5 ms per-test total (37%), larger than the test body, both hooks, mock teardown and result encoding combined.

I hit this on my own machine: 4,242 files, 45 MB, all from runs killed mid-flight during a long session.

Proposal

Stat one known path instead of listing the directory. temp_file/temp_dir write a <id>_.mark marker (a redirect, not a fork); the trap returns early when it is absent, which is the common case since most tests create no temp file. The marker shares the <id>_ prefix, so the existing rm removes it along with everything else.

Cost becomes O(1) in the size of the directory:

leftovers	before	after
0	498 ms	520 ms
2000	635 ms	520 ms
5000	978 ms	542 ms

Behaviour change

A file a test writes into BASHUNIT_TEMP_DIR by hand, named with its own test id, is no longer removed — only files handed out by temp_file/temp_dir are. Finding a hand-rolled one is exactly the directory scan being removed.

[Chemaclass]

Done in #1270 (merged).