[wasikarn]

Closes the two test gaps that the third fresh-context review of #632 left open. This PR is test-only: scripts/_lib/stray-procs.sh doesn't change. Refs #158.

Why

• set -e load path: stray_procs_report promises it never returns non-zero. But the only set -e row covered the ps-failure path, so dropping || load="" left 21/21 green. A bash -e caller that found strays while the load reader failed would then have exited 1.
• Real ps call: every row stubbed the private _sp_ps/_sp_load, so the real ps -U <uid> -o pid=,etime=,%cpu=,command= never ran. Changing etime= to etimes=, which macOS ps doesn't have, still passed.

What

tests/scripts/test-gauntlet-stray-procs.sh gets 4 more rows (21 to 25):

• A bash -ec caller with strays and a failing _sp_load reaches the next command.
• PATH shims, with the real readers running:
  • ps shim: behaves like macOS ps wherever the report depends on it. An unknown -o keyword is an error, and columns come out in the order -o asks for. With no -o list, it prints the default PID TTY TIME CMD columns under a header. It lists rows only for -U <your uid>. Checks that the real call feeds the report.
  • uptime shim: prints the macOS format (load averages: 3.21 2.00 1.00) and the Linux format (load average: 4.56, 2.00, 1.00). Checks that both parse.
• The shim dir comes from tests/_lib/harness.sh (fresh_tmpdir, with _cleanup_trash on EXIT).

No version bump; merge-pr.sh does that at merge time.

Verified

• 25/25 pass, and shellcheck (warning level) is clean.

• Mutations of stray-procs.sh: each of these 7 makes the suite fail. The last two came from the first review of this PR, and the lib was restored after each (git diff empty):

  Mutation	Failing rows
  drop || load=""	1
  etime= → etimes=	3
  ps -U "$(id -u)" → ps -A	3
  load pattern made macOS-only	1
  load split on spaces only	1
  -o columns reordered	1 for etime=,pid=,…; 3 for orders that move %cpu or command, such as pid=,%cpu=,etime=,command=
  -o list left out	3

• audit.sh exits 0, and the pre-push gauntlet passed all layers on cce630fd.

[wasikarn]

Fresh-context review: FIX FIRST @ cce630f

Two-axis review (mattpocock-skills:code-review). A separate subagent ran each axis against merge-base a14ec7e.

Spec (MISSING=0 CREEP=0 WRONG=1)

• Both gaps are closed, confirmed by mutation:
  • dropping || load="" fails 1 row
  • etimes= fails 3
  • renaming _sp_ps breaks 11 stub rows, while all 3 shim rows stay green, so they survive a helper rename.
• Wrong: the PR body says the ps shim "behaves like macOS ps". But it only checks the column names. Deleting -o … or reordering the columns stays 25/25 green.

Standards (HARD=1 JUDGEMENT=3)

• Hard (test-honesty pr-test-analyzer: add 'weakened-to-pass / gate-gaming' critical-gap class #5): the "real ps call" row's name claims more than it checks. Swapping to -o etime=,pid=,… stays green, though on a real machine awk would then read the wrong fields.
• Judgement:
  • The new inline bash -c cases repeat a shape that run() can't take. Keeping them: a helper adds more than it saves.
  • harness.sh is sourced mid-file rather than at the top. Will fix.
  • SHIM=$(fresh_tmpdir) has no guard. Will fix.

Fix: the shim will print columns in the requested order, and without -o it prints macOS's default columns. Reordering or dropping -o then breaks the report the way it would on a real machine.

[wasikarn]

Fresh-context review: MERGE @ 5b80d35

Second two-axis review (mattpocock-skills:code-review). New subagents ran it against merge-base a14ec7e.

Spec (MISSING=0 CREEP=0 WRONG=1). Both gaps are closed, confirmed by mutation on a git archive copy:

• dropping || load="" fails 1 row
• etimes=, ps -A and leaving out the -o list each fail 3
• every -o reorder tried fails 1 to 3
• renaming _sp_ps breaks the 11 stub rows while all 3 shim rows stay green

The one "wrong" item was the PR table's reorder count, which depends on the order chosen. It's now corrected in the body.

Standards (HARD=0 JUDGEMENT=4). Clean on:

• the harness contract (source plus its own EXIT trap, a guarded shim dir)
• no rm or direct trash
• no home paths or bare grep
• mocks at the process boundary, not on the unit

Judgement calls, none blocking:

• the uptime rows fail alongside a broken ps, since the report returns early with no strays
• the unquoted shim heredoc needs many \$ escapes
• the shim run line is repeated twice
• the header says "last rows", but the wiring row comes after them

Verdict: MERGE. No hard violation, and no spec gap in the code.