diff --git a/docs/superpowers/specs/2026-06-20-perf-comparison-suite-design.md b/docs/superpowers/specs/2026-06-20-perf-comparison-suite-design.md
new file mode 100644
index 00000000..db2b29c0
--- /dev/null
+++ b/docs/superpowers/specs/2026-06-20-perf-comparison-suite-design.md
@@ -0,0 +1,285 @@
+# Performance comparison suite (shell vs Go distrobox)
+
+## Goal
+
+Build a developer-local performance test suite that measures wall time, peak
+memory, and CPU counters of a `distrobox` executable across a fixed set of
+scenarios. The suite takes the executable path as an argument so the same
+scenarios can be run against either the shell implementation (`main` branch's
+`./distrobox`) or the Go rewrite (`next` branch's `./bin/distrobox`), producing
+result directories that a separate comparison tool diffs into a markdown
+report.
+
+Out of scope: functional/behavioral testing (covered by `tests/compare.sh`),
+CI integration, multi-host comparison, flamegraph rendering, Go pprof
+profiling.
+
+## Constraints / decisions already made
+
+- **Scope:** pure-overhead scenarios AND end-to-end container lifecycle scenarios.
+- **Profiling depth:** black-box (hyperfine, `/usr/bin/time -v`, strace optional)
+  plus `perf stat` / `perf record` on both implementations. No language-specific
+  profilers (no pprof, no `bash -x`).
+- **Container engine:** podman only, pinned via `DBX_CONTAINER_MANAGER=podman`.
+- **Test image:** `docker.io/library/alpine:latest`. Pre-pulled once, never
+  removed by the suite.
+- **Harness language:** POSIX shell (`/bin/sh`), mirroring the existing
+  `tests/compare.sh` style. Tools assumed present: `hyperfine`, `jq`,
+  `/usr/bin/time` (GNU), `perf`, `podman`.
+
+## Repository layout
+
+```
+bench/
+├── run.sh            # entry point: ./bench/run.sh <executable> [opts]
+├── compare.sh        # ./bench/compare.sh <result-dir-A> <result-dir-B>
+├── lib/
+│   ├── hyperfine.sh  # hyperfine wrapper, JSON export
+│   ├── time.sh       # /usr/bin/time -v wrapper + parser to JSON
+│   ├── perf.sh       # perf stat + perf record wrappers
+│   ├── container.sh  # podman setup/cleanup, tracker file ops
+│   └── common.sh     # logging, scenario loader, preflight
+├── scenarios/
+│   ├── 01-startup-help.sh
+│   ├── 02-startup-version.sh
+│   ├── 03-subcommand-help.sh
+│   ├── 04-assemble-parse.sh    # may be a no-op if --dry-run not supported; see §Open
+│   ├── 10-list-empty.sh
+│   ├── 11-list-many.sh
+│   ├── 20-create-rm.sh
+│   ├── 21-enter-exec.sh
+│   └── 22-ephemeral-true.sh
+├── fixtures/
+│   └── assemble.ini
+└── results/
+    ├── <label>/<run-id>/...
+    └── comparisons/<labelA>-vs-<labelB>-<timestamp>.md
+```
+
+`bench/results/` is gitignored.
+
+## Scenarios
+
+Each scenario file under `bench/scenarios/` defines three shell functions:
+
+- `scenario_setup` — runs once before any measurement; can pre-create containers
+  via `container_create_tracked` (defined in `lib/container.sh`)
+- `scenario_command` — the command whose execution is measured. Echoed to
+  stdout as a single line that `hyperfine` / `time` / `perf` consume.
+- `scenario_cleanup` — runs once after all measurements; typically calls
+  `container_cleanup_tracked`
+
+Scenarios may also export `SCENARIO_WARMUP`, `SCENARIO_RUNS`, and
+`SCENARIO_DESCRIPTION` variables read by the harness.
+
+### Pure overhead
+
+| # | Name | Command | What it isolates |
+|---|------|---------|------------------|
+| 01 | `startup-help` | `$EXEC --help` | Binary load + arg parser cost |
+| 02 | `startup-version` | `$EXEC --version` | Same cost class, distinct codepath |
+| 03 | `subcommand-help` | `$EXEC create --help` | Cost of dispatching into a subcommand |
+| 04 | `assemble-parse` | `$EXEC assemble create --file fixtures/assemble.ini --dry-run` (if supported, else dropped — see §Open questions) | Pure `.ini` parsing cost |
+
+Iteration target: `--warmup 3 --min-runs 50`.
+
+### Engine-touching, no container created
+
+| # | Name | Command | What it isolates |
+|---|------|---------|------------------|
+| 10 | `list-empty` | `$EXEC list` with zero `dbx-bench-*` containers present | Cost of engine-query path on empty result |
+| 11 | `list-many` | `$EXEC list` with 20 pre-existing dummy distrobox-labelled containers (created by the harness via `podman create` directly to keep setup fast) | Cost of formatting/filtering many rows |
+
+Iteration target: `--warmup 3 --min-runs 30`.
+
+### Full lifecycle
+
+| # | Name | Per-iteration | What it measures |
+|---|------|---------------|------------------|
+| 20 | `create-rm` | `create --image alpine --name $NAME` then `rm --force $NAME` | Full create+destroy cost |
+| 21 | `enter-exec` | `enter --name bench -- /bin/true` against a container pre-created in setup | Per-exec overhead (most user-visible metric) |
+| 22 | `ephemeral-true` | `ephemeral --image alpine -- /bin/true` | Full ephemeral lifecycle |
+
+Iteration target: `--warmup 1 --runs 10` (per-iteration cost too high for 50).
+
+## Measurement layers
+
+Each scenario is invoked three times, once per layer, in this order, so layers
+don't interfere with each other (notably `perf` can perturb timing):
+
+| Layer | Tool & flags | Output | Captures |
+|-------|--------------|--------|----------|
+| Wall time | `hyperfine --export-json <out> --warmup N --runs M -- '<cmd>'` | `<scenario>.hyperfine.json` | mean, stddev, median, min, max wall clock |
+| Resource accounting | `/usr/bin/time -v -o <out> <cmd>` (single run; see note) | `<scenario>.time.txt` (raw) + `.time.json` (parsed by `lib/time.sh`) | peak RSS, user/sys CPU, page faults, voluntary/involuntary context switches, FS inputs/outputs |
+| CPU counters | `perf stat -x, -o <out> -e task-clock,cycles,instructions,cache-references,cache-misses,page-faults,context-switches <cmd>` (single run; see note) | `<scenario>.perf-stat.csv` | total instructions, IPC, cache-miss rate, syscall pressure proxies |
+| CPU sampling (opt-in via `--profile`) | `perf record -F 99 -g -o <out> <cmd>` | `<scenario>.perf.data` (large, gitignored) | flamegraph-able call samples |
+
+**Why single-run for time/perf-stat:** hyperfine already characterises
+wall-time variance across many runs. Peak RSS, instructions, cache-miss rate,
+and the other counters are dominated by deterministic per-invocation costs and
+are typically stable across runs; the extra runs aren't worth the wall-clock
+budget. If a scenario reveals instability, it can be re-run by hand under
+`time -v` / `perf stat` to confirm.
+
+Caveats called out in the suite's README and printed at run start:
+
+- **`perf_event_paranoid`**: `perf stat` and user-space `perf record` work
+  without sudo when the value is ≤ 2. The harness reads `/proc/sys/kernel/perf_event_paranoid`
+  on start and warns (but proceeds without the perf layer) if it's higher.
+- **Shell pipelines & `/usr/bin/time -v`**: peak RSS reports only the parent's
+  RSS, not children. This understates absolute memory for the shell
+  implementation but is consistent across runs, so ratios remain meaningful.
+- **Shell perf-stat noise**: every `bash` invocation forks many short-lived
+  utilities; the counters reflect that, which is part of what we're measuring,
+  not a flaw.
+
+## Container & engine integration
+
+- All scenarios run with `DBX_CONTAINER_MANAGER=podman` exported by the harness
+  so distrobox doesn't auto-detect.
+- The harness pulls `docker.io/library/alpine:latest` once at suite start and
+  asserts it's present; teardown does not remove it.
+- **Run ID**: `RUN_ID="$(date +%s)-$$"` — timestamp plus PID, unique per run
+  even if two run concurrently.
+- **Naming**: containers created by the harness are named
+  `dbx-bench-<run-id>-<scenario>-<iteration>`.
+- **Tracker file**: `bench/results/<label>/<run-id>/containers.list`. Every
+  name is appended the instant it's created (whether via `distrobox create` or
+  the direct `podman create` shortcut used by `list-many` setup). Cleanup
+  iterates the tracker and removes by **exact** name with `podman rm -f <name>`.
+  Never relies on `podman ... --filter name=` (not a partial-match contract
+  the harness should depend on for destructive ops).
+- **Belt-and-suspenders label**: containers also get
+  `--label distrobox.bench.run=<run-id>`. If the harness crashes and the
+  tracker is incomplete, `podman ps -a --filter label=distrobox.bench.run=<run-id> -q | xargs -r podman rm -f`
+  cleans up.
+- **Pre-flight orphan scan**: at run start, look for any container with the
+  `distrobox.bench.run` label from *any* previous run. If any exist, the
+  harness aborts with a clear message; pass `--clean-orphans` to remove them
+  first. The harness never silently destroys state it didn't create in the
+  current invocation.
+- **`list-many` setup detail**: 20 dummy containers are created via
+  `podman create --label manager=distrobox --label distrobox.bench.run=<run-id>
+  --name dbx-bench-<run-id>-list-many-NN alpine /bin/true`. This is fast (no
+  start, no init-script) and is sufficient for `distrobox list` to find them.
+  *Risk*: if the label distrobox actually uses for detection differs from
+  `manager=distrobox`, the count won't match. The harness verifies the count
+  before running the scenario; if it doesn't match, the scenario is marked
+  skipped with a diagnostic. The exact label/filter contract is read from
+  `pkg/commands/list.go` and `distrobox-list` during implementation.
+
+## Output format
+
+### Per-run directory
+
+```
+bench/results/<label>/<run-id>/
+├── meta.json
+├── containers.list
+├── scenarios/
+│   ├── 01-startup-help.hyperfine.json
+│   ├── 01-startup-help.time.txt
+│   ├── 01-startup-help.time.json
+│   ├── 01-startup-help.perf-stat.csv
+│   └── (one set per scenario, plus .perf.data if --profile)
+└── summary.md
+```
+
+`meta.json` shape:
+
+```json
+{
+  "label": "next",
+  "run_id": "1718901234-12345",
+  "started": "2026-06-20T12:30:00Z",
+  "finished": "2026-06-20T12:33:21Z",
+  "executable": "/abs/path/to/bin/distrobox",
+  "executable_sha256": "…",
+  "executable_size_bytes": 12742149,
+  "host": {
+    "kernel": "6.19.10-1-default",
+    "cpu_model": "…",
+    "podman_version": "5.8.1",
+    "perf_event_paranoid": 2
+  },
+  "scenarios_run": ["startup-help", "startup-version", "..."],
+  "scenarios_skipped": [{"name": "assemble-parse", "reason": "--dry-run not supported by this binary"}]
+}
+```
+
+`summary.md` is a single markdown table: scenario × (mean ± stddev, peak RSS,
+instructions). Generated unconditionally at the end of `run.sh`.
+
+### Comparison directory
+
+`bench/results/comparisons/<labelA>-vs-<labelB>-<timestamp>.md` produced by
+`compare.sh`. Header section:
+
+```
+Binary A (label: next):  /abs/path  sha256:…  size: 12,742,149 bytes
+Binary B (label: main):  /abs/path  sha256:…  size: …
+Podman: 5.8.1 (matches)
+Host kernel: 6.19.10-1-default (matches)
+Run-id A: 1718901234-12345
+Run-id B: 1718900000-12000
+```
+
+Per-scenario rows: mean wall time A, B, ratio (B/A), marker
+(`Δ < stddev` = "within noise"; otherwise `Nx faster/slower`); peak RSS A, B,
+ratio; instructions A, B, ratio.
+
+Refuses to compare if podman versions differ. Pass `--allow-engine-drift` to
+override (the difference is annotated in the report header).
+
+## Invocation
+
+```
+./bench/run.sh <executable> [--label NAME] [--scenarios LIST]
+                            [--profile] [--clean-orphans]
+```
+
+- `<executable>` — positional, required, absolute or relative path. The harness
+  resolves it to an absolute path before running.
+- `--label NAME` — directory name under `bench/results/`. Defaults to the
+  executable basename (so `./bin/distrobox` becomes `distrobox`). Multiple runs
+  with the same label coexist as sibling `<run-id>` subdirs; nothing is ever
+  deleted by `run.sh`.
+- `--scenarios LIST` — comma-separated names or shell glob (`startup-*`,
+  `2*`). Default = all.
+- `--profile` — enable `perf record` per scenario. Off by default
+  (`.perf.data` files are large).
+- `--clean-orphans` — only honoured at run start: removes containers from
+  previous runs (matched by the `distrobox.bench.run` label) before pre-flight.
+
+```
+./bench/compare.sh <result-dir-A> <result-dir-B> [--allow-engine-drift]
+```
+
+## Risks / open questions
+
+1. **`distrobox assemble create --dry-run`** — needs verification that this
+   subcommand/flag exists on both implementations. If it doesn't, scenario 04
+   is dropped and noted in the spec. Read `distrobox-assemble` and
+   `pkg/commands/assemble.go` during implementation Phase 1.
+2. **Label distrobox uses to identify its containers** — `list-many`'s direct
+   `podman create` shortcut depends on the exact label/filter contract. Verify
+   during implementation Phase 1 by reading `distrobox-list` and
+   `pkg/commands/list.go`. If the contract is more involved (e.g., requires
+   specific label *value* per-distrobox), the shortcut is replaced with a
+   parallel `distrobox create --yes` loop, accepting the slower setup cost.
+3. **`--version` flag collision** — `next`'s help shows `--version, -v` AND
+   `--verbose, -v` sharing `-v`. The harness always uses long forms
+   (`--version`, `--verbose`) to avoid ambiguity across implementations.
+4. **Time/perf overhead of `enter-exec`** — `enter` may include
+   significant podman-exec time that dominates the implementation difference.
+   This is acceptable; the suite reports what it reports. The spec is honest
+   about this in the README.
+
+## Non-goals (reaffirmed)
+
+- No CI wiring.
+- No flamegraph generation (raw `perf.data` only).
+- No Go pprof harness — if needed, a sibling `bench/pprof.sh`, not a feature
+  of this suite.
+- No automatic statistical significance testing beyond hyperfine's stddev /
+  the "Δ < stddev → noise" marker.