diff --git a/.claude/rules/perf-fork-budget.md b/.claude/rules/perf-fork-budget.md
new file mode 100644
index 00000000..03918d62
--- /dev/null
+++ b/.claude/rules/perf-fork-budget.md
@@ -0,0 +1,80 @@
+---
+paths:
+  - "src/**/*.sh"
+  - "tests/acceptance/bashunit_*forks*_test.sh"
+---
+
+# Performance: Fork Budget & Census Method
+
+bashunit's dominant runtime cost on Bash 3.2 (macOS) is **process forks**, not
+shell execution. A fork costs ~1-3ms; the acceptance suite spawns ~258 nested
+`./bashunit` runs, so one avoidable fork per file or per test multiplies fast.
+PRs #801-#811 removed the per-test and cold-start forks; this file records how
+to measure and the traps found, so future work starts from evidence.
+
+## Measuring: shim census is ground truth, traces overcount
+
+**PATH-shim census (authoritative).** Wrap a binary, count invocations:
+
+```bash
+real_awk="$(command -v awk)"          # resolve BEFORE prepending the shim dir
+printf '#!/usr/bin/env bash\necho x >> /tmp/count\nexec "%s" "$@"\n' "$real_awk" > shim/awk
+chmod +x shim/awk
+PATH="shim:$PATH" ./bashunit --no-parallel fixture_test.sh
+grep -c . /tmp/count
+```
+
+Caveats: binaries pinned at startup via `command -v` (`$GREP`, `$MKTEMP`,
+`$CAT`) bypass PATH shims — trace those instead; and shims are unreliable on
+Git Bash (skip such tests on Windows).
+
+**`bash -x` trace census (cheap but inflated).** `PS4='+ ' bash -x ./bashunit …`
+also counts trace lines **re-echoed inside captured test output**, so it can
+overcount 10-20x (one real `grep` appeared 24 times). Use it to *locate* fork
+sites (`PS4='+X ${BASH_SOURCE##*/}:$LINENO> '`), never to *count* them. It also
+misses forks inside `--parallel` workers. Anchor real execs with
+`grep -cE '^\++ +/[^ ]*mktemp$'`-style patterns (absolute path, exact argv
+shape) to skip `command -v` probes and variable assignments.
+
+**Regression tests.** Budgets are enforced in
+`tests/acceptance/bashunit_coldstart_forks_test.sh` (cold start: mktemp/perl/
+grep/mkdir) and `tests/acceptance/bashunit_run_forks_test.sh` (run path:
+grep/cat/sed/sort/awk budget, run-dir cleanup). Extend these when you remove a
+fork class; they are the RED test of the TDD cycle.
+
+## Replacement patterns that worked
+
+| Fork | Replacement | Example |
+|------|-------------|---------|
+| `echo x \| grep -c fixed` | `case` glob match | #803, #804 |
+| `cat <<EOF` heredoc emit | `printf '%s\n' "$payload"` | #806 (was 1 fork/test) |
+| `sed -n Np` per line in a loop | one `while IFS= read -r` pass over the file | #807 (quadratic → flat) |
+| `declare -F \| awk '{print $3}'` | `compgen -A function` builtin (identical output) | #810 |
+| small-list `awk \| sort \| awk` | pure-bash filter + insertion sort | #809 (tens of items only) |
+| N × `mktemp` scratch files | one run-unique dir + fixed names, lazy `>>` creation | #801, #811 |
+| probe fork + first-use fork | probe does the real work and seeds the return slot | #802 (clock perl) |
+
+`shopt -s extdebug` for `declare -F`: enable it **inside the capture subshell
+only** — toggling it in the caller's shell clobbers caller state (#808).
+
+## Where pure bash LOSES — measured, do not "optimize" these
+
+- **`… | awk` over ≳100 lines**: a `while read` loop was 5x slower than the awk
+  fork (19ms vs 3.5ms for 600 `declare -F` lines). File scans (provider map,
+  duplicate check) stay awk.
+- **`${var//pattern/}` on large strings**: quadratic on Bash 3.2 — 2.7 s where
+  awk takes 3.5ms. Never string-replace over big captures.
+- **Regex assertions** (`assert_matches`): Bash 3.0 has no `[[ =~ ]]`; the
+  `grep -E` fork is mandatory.
+- **Single-file build artifact**: sourcing `bin/bashunit` is *not* faster than
+  sourcing `src/*.sh` (parse time dominates, file opens don't).
+- **`tput cols` at startup**: returns 80 on non-tty; snapshots depend on that
+  width. Not removable.
+
+## Current budgets (plain 1-test file run, Bash 3.2 macOS)
+
+~3 `awk` (provider map ×2 — the counting subshell can't share its cache with
+the runner — plus the duplicate check), 1 `perl` ×2 clock reads (start/end;
+no `EPOCHREALTIME` before Bash 5), 1 `base64` capability probe, 1 `mkdir`,
+1 `tput`. Per-test cost is fork-free. Cold start ~50ms, ~31ms of which is
+sourcing `src/` (irreducible without lazy-loading, rejected in #798).