v0.43.1

Fixed

• Fixed #906: compile_hooks.py's _DYNAMIC_CALL_RE dynamic-dispatch detector
  matched && "$name" even when "$name" was a test operand inside
  [[ ... ]]/[ ... ], not a real call through a variable -- the exact shape
  that grew a compiled hook from ~122 KB to 135,889 B during #899's size work,
  with no visible symptom besides the size. Operands inside bracket-test
  spans are now masked before the dynamic-dispatch scan runs, and when
  tree-shaking genuinely is disabled, compile_hooks.py --apply now prints
  the reason and the triggering line to stderr instead of staying silent.

• Corrects a process-count claim from this range's own v0.43.0 entries (#1011, #1017): "1 (tail only)" / "a second spawn, wc" counted only exec'd binaries -- programs named on PATH -- never the bash subshell forks a $(...) command substitution or a | pipe also creates, even when nothing inside them execs a new binary at all. A round-2 follow-up to this same fragment then asserted the regression it found against the wrong number: the whole below-threshold hook invocation's own real fork count (measured via strace -f -e trace=clone,clone3,fork,vfork on CI, every clone/fork/vfork anywhere in the process tree -- libs sourced, tool-detection probes, the background save detach, all of it) checked against 6, a figure that only ever described _pt_scan_lines's own local contribution, not the whole hook. CI measured 67 on every run across every Linux leg, not below 6.

  Measured the right way now: tests/test_post_tool_total_processes_1011.py runs the identical strace harness against three real trees on CI -- the tagged v0.42.4 release, the tagged v0.43.0 release immediately preceding this one, and this fix -- instead of asserting against a hand-picked threshold. CI-measured whole-hook totals for the below-threshold no-save path, minimum of 9 runs each, identical on every Linux leg: v0.42.4 = 68, v0.43.0 = 69, this fix (HEAD) = 67. The fix itself is real and small -- _pt_scan_lines now sets globals instead of running inside an outer command substitution, and its newline count uses a herestring into wc -l instead of a printf | wc pipe, removing two subshell forks -- and it does land below both baselines, but the whole-hook invocation still creates on the order of 67 processes for one no-save PostToolUse call. On the Windows 11 25H2 build #1011 originally reported against, each of those is a real kernel Token that leaks until reboot; a two-process improvement does not change that order of magnitude. Attributing which script lines and sourced libraries account for the bulk of the 67 is tracked separately, not in this fragment: see the newly filed follow-up issue.

  The cooldown-active path's own reduction (skips the scan entirely) is unaffected by this correction and remains genuine.

• Fixed #1014: a hung pytest leg on CI used to be cancelled by its own
  job-level timeout-minutes with no log at all, leaving no way to tell
  which test hung. pyproject.toml now sets faulthandler_timeout = 900
  so a stuck test dumps every thread's traceback well before CI's own
  25-minute step timeout fires, the "Run tests" step now runs under its
  own timeout-minutes: 25 and tees output to a log file that is uploaded
  as an artifact under if: always(), and a Windows leg additionally
  lists any surviving wscript.exe/bash.exe/conhost.exe/
  powershell.exe processes at teardown. Proven on Linux (run 38122057767,
  job 114445674357): the faulthandler traceback naming the stuck test is
  present in the kept log. The equivalent Windows failure mode (a job
  ending cancelled with no log at all, run 38122057767, job
  114418645594) is not shown fixed by this change; tracked separately.

• #1029: three unconditional per-call forks removed from post-tool-hook.sh, on top of CI-measured strace totals this range already tracks (v0.42.4 = 68, v0.43.0 = 69, prior HEAD per #1028 = 67). _stdin_json_string (cwd, session_id, transcript_path) was called through X=$(...) command substitution at three sites above the fast/slow-path split, each forking a subshell even though the function itself is pure parameter expansion -- the same cost class lib-clock.sh's own _remember_date_into fix already removed elsewhere in this hook. Ported to the fork-free _stdin_json_string_into form session-start-hook.sh already used (#665, part of #660): identical matching algorithm (same first-occurrence scan, same #344 gap, same #829 double-backslash collapse), no behaviour change, three fewer forks on every invocation of this hook, cold-start and warm path alike.

  tests/test_post_tool_total_processes_1011.py's own fixture measures the cold-start path only (a brand-new project with no warmed caches), which is why #1011/#1028's own strace totals land near 67-69 rather than the much lower warm-path count a real session's later tool calls pay once a project has run once. On that warm path, the only forks left after this cut are _pt_scan_lines's own tail+wc pair, already documented in-code as unavoidable without reintroducing #1017's quadratic bash-expansion regression -- every other candidate this issue named (bootstrap-dirs.sh's directory creation, the jq config read, the mktemp+mv atomic-write patterns) traced to cold-path-only work already gated behind existence checks (#230/#666) or work inside the byte-pinned lib-memory-dir.sh (tests/test_case_divergence_298.py forbids touching it).

  tests/test_post_tool_total_processes_1011.py now asserts an absolute budget (strictly below 67, this round's own prior total) for both the below-threshold and cooldown paths, in addition to the existing relative tag-comparisons, so a regression is caught even on a no-network CI leg where neither tagged baseline tree can be materialized. Confirmed by this pull request's own CI run, the same strace harness against the same three trees: the whole-hook, cold-start, below-threshold total is v0.42.4 = 68, v0.43.0 = 69, HEAD (this fix) = 64 -- three fewer than the 67 reasoned above, matching the three forks this fix removes. The warm-call cost (a project that already ran once, caches populated) is not measured by this fixture and no number is claimed for it.