perf: the setup pass runs its stages in subshells, not a bash each

CHANGELOG.md

@@ -7,6 +7,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- **A cold launch's setup pass starts one shell, not one per stage.** Each stage (session
+  name, zellij, codex, title, onboarding, trust) ran in a nested `bash -c`, and every new
+  bash reads the image's `$BASH_ENV`. An image that points it at a heavy shell profile
+  (for example a ROS environment) can pay about 1.1s a read inside `devpod ssh`, so the
+  trip took about 11s. The stages now run in subshells of the pass
+  (`( eval '<script>' )`), which keep each stage's `exit`, `set -u` and parse errors to
+  itself as the nested bash did. Measured over six cold launches of such an image, the
+  setup trip went from 10.9s to 11.1s down to 6.7s to 6.9s.
+
 ## [0.60.0] - 2026-10-16
 
 ### Added

docs/performance.md

@@ -120,6 +120,25 @@ failing. A master that has gone away leaves a socket the next client unlinks, an
 resident `devpod ssh --stdio` process and a `docker exec` per key, and `dl` must
 not be the reason a container never goes idle.
 
+## One shell for the setup pass
+
+A cold launch sends one `devpod ssh --command` trip that runs dl's setup stages
+and then the tools probe. Each stage runs in a subshell of that trip's shell, not
+in a `bash` of its own. The reason is `$BASH_ENV`: every new non-interactive bash
+reads it, and an image may point it at a heavy shell profile (for example a ROS
+environment). On the image measured below, one read inside a `devpod ssh`
+session costs about 1.1s.
+
+Measured on 2026-10-12, on one host with the local docker provider and devpod
+0.26.3, over six cold launches of a large ROS repository whose image sets
+`$BASH_ENV`, on six fresh branches. With a `bash -c` per stage the trip took
+10.90s to 11.14s. With subshells it took 6.66s to 6.90s, so about 4.2s comes
+off every cold launch of that repo. The saving on an image with no `$BASH_ENV` is not measured. It should
+be the cost of starting a bash per stage, which is small. A warm launch makes no setup trip, so it is not affected.
+
+`the_pass_reads_the_images_bash_env_once_however_many_stages_it_runs` and
+`dl/tests/setup_pass_shells.rs` pin the count at one read per pass.
+
 ## Measuring launch time
 
 Set `DEVLAUNCH_TIMING=1` and a `dl` command ends with one summary on stderr,