diff --git a/.agnostic-ai/AGNOSTIC_AI.md b/.agnostic-ai/AGNOSTIC_AI.md
index 1c985231..20695bcf 100644
--- a/.agnostic-ai/AGNOSTIC_AI.md
+++ b/.agnostic-ai/AGNOSTIC_AI.md
@@ -13,8 +13,8 @@ Preserve Bash 3.0 parsing and public APIs; follow sibling tests/code. Update doc
 Main selects checks by changed behavior/risk and reuses unchanged passes. Small fixes: affected tests, ShellCheck, `make lint`; strict mode for options, errors, hooks, or workers. Cross-cutting runtime changes/releases need:
 
 ```bash
-make test
-make test/strict
+BASHUNIT_PARALLEL_JOBS=4 make test
+BASHUNIT_PARALLEL_JOBS=4 make test/strict
 make sa
 make lint
 ```
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 62430297..5026d9bd 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,35 +8,35 @@
 
 ### Fixed
 
+#### Assertions and doubles
+
 - Date assertions reject trailing text on macOS and parse colon-separated timezone offsets consistently across platforms (#1404)
+- `assert_string_not_starts_with` joins every haystack argument and uses the test name as its failure label, matching its siblings (#1403). `assert_matches` and `assert_not_matches` accept patterns starting with `-` and fail when grep rejects a pattern (#1398)
+- `assert_files_equals` and `assert_files_not_equals` fail and name the missing path when either file does not exist (#1397)
 - Snapshots record and update literal `-n`, `-e`, `-E` and combined echo flag values correctly (#1402)
-- Generated workflows run test directories containing whitespace as a single path without requiring a committed `.env` (#1425)
-- Annotations, tags, header counts and duplicate checks support functions declared without parentheses (#1422)
+- Spies keep multiline arguments in one recorded call, including comparisons and failure logs (#1399). Call argument assertions fail when the spy was never called or the requested call index does not exist (#1400)
+- `bashunit::unmock` restores shell functions replaced by spies, mocks or mock sequences (#1401)
+
+#### Setup and selection
+
+- `install.sh` and `upgrade` keep the installed binary when a download fails or is empty; installation also keeps it when checksum verification fails (#1406, #1407)
+- `install.sh` resolves relative destinations from the caller's directory, reports the resolved path (#1408), and rejects prefixed or incomplete version numbers instead of treating them as directories (#1409)
+- `init` quotes the bootstrap path and custom tests directory in `.env`, so paths containing spaces work on later runs (#1411). Generated workflows pass a tests directory containing whitespace as one path without requiring a committed `.env` (#1425)
+- `.bashunitrc` suite options stay out of the environment; global settings before and after suite sections still apply (#1410)
+- Include, inline and exclusion filters preserve `test_` inside names such as `latest_version` (#1414). `--debug` keeps an existing test file or directory as the test path instead of taking it as the trace file (#1405)
+
+#### Test discovery and execution
+
+- Annotations, tags, header counts and duplicate checks support functions declared without parentheses (#1422). Duplicate definitions with an opening brace on the next line are reported instead of replacing the first definition (#1421)
 - Echo data providers keep `#` and parentheses, and report malformed rows without dropping later rows. Quoted values keep carriage returns on Windows (#1419)
-- A green file or filtered subset preserves rerun-cache entries for tests it did not execute (#1418)
-- SIGTERM and SIGHUP run teardown hooks, stop parallel workers, and remove the run's scratch files before exiting with status 1 (#1417)
-- Sequential stop-on-failure runs write requested reports and the rerun cache for the tests that ran (#1415)
-- Include, inline, and exclusion filters preserve `test_` inside names such as `latest_version` (#1414)
-- `.bashunitrc` suite options stay out of the environment, while global settings before and after suite sections still apply (#1410)
-- `install.sh` rejects prefixed or incomplete version numbers instead of treating them as directories (#1409)
-- `install.sh` resolves relative destinations from the caller's directory and reports the resolved installation path (#1408)
-- `bashunit::unmock` restores a shell function replaced by a spy, mock, or mock sequence (#1401)
-- Spies keep a multiline argument in one recorded call, including call comparisons and failure logs (#1399)
-- Parallel results and hook output stay separate for test paths that differ by slashes and underscores (#1416)
-- Coverage report paths are checked before tests run, and LCOV or HTML write failures make the run fail instead of reporting success (#1412)
-- `upgrade` keeps the installed binary when a download fails or returns an empty file (#1407)
-- `install.sh` keeps the installed binary when a download fails, is empty, or fails checksum verification (#1406)
-- `--debug` leaves an existing test file or directory as the test path instead of taking it as the trace file (#1405)
-- Call argument assertions fail when the spy was never called or the requested call index does not exist (#1400)
-- `assert_matches` and `assert_not_matches` accept patterns starting with `-` and fail when grep rejects a pattern (#1398)
-- `assert_files_equals` and `assert_files_not_equals` fail and name the missing path when either file does not exist (#1397)
-- TAP reports keep test titles containing line breaks on a single result line (#1413)
+- SIGTERM and SIGHUP run teardown hooks, stop parallel workers and remove scratch files before exiting with status 1 (#1417)
+- Parallel results and hook output stay separate for paths that differ by slashes and underscores (#1416). Relative hidden paths are counted and keep separate results from matching visible paths (#1393)
+
+#### Reports and reruns
 
-- `init` writes the bootstrap path and a custom tests directory to `.env` quoted, so a directory with a space runs afterwards instead of breaking every next run (#1411)
+- Sequential stop-on-failure runs write requested reports and the rerun cache for tests that ran (#1415). A green file or filtered subset preserves cache entries for tests it did not execute (#1418)
+- Coverage report paths are checked before tests run; LCOV or HTML write failures make the run fail (#1412). TAP reports keep test titles containing line breaks on one result line (#1413)
 - Failure source context stays inside a one-line test instead of listing the next function's assertions (#1420)
-- Parallel runs count tests under relative hidden paths and keep their results separate from matching visible paths (#1393)
-- Duplicate test functions are detected when the opening brace is on the next line instead of silently replacing the first definition (#1421)
-- `assert_string_not_starts_with` treats variadic haystack arguments as one newline-joined value and keeps failures labeled with the test's name (#1403)
 
 ### Changed