AGENTS.md

@@ -238,12 +238,13 @@ The leg re-checks the conversions and nothing else — each fixture's emitted ma
 document, its error code — leaving the corpus's pairing, uniqueness, source positions and
 byte-level equality to the Node suite that owns them.
 
-Every leg announces its name and its image before it runs and its elapsed time after, `publish.sh`
-alongside `ci.sh`, so a long run reads as progress rather than as a hang — which is what a silent
-one cost the `0.1.0` release. A leg added later owes the same marker, and a leg that buffers its
-output to read something out of it streams and keeps the copy in a file: `tee /dev/stderr` reopens
-the stream, so under the `> log 2>&1` a reader runs locally the two offsets advance independently
-and punch NUL holes through each other's lines (4d).
+Every leg announces its name, and the image where it runs in one, before it runs and its elapsed
+time after, `publish.sh` alongside `ci.sh`, so a long run reads as progress rather than as a hang.
+A leg added later owes the same marker, and a function a leg runs chains its statements with `&&`,
+because the `||` that captures the leg's status suspends `set -e` for everything it calls. A leg
+whose output is both streamed and grepped keeps the copy in a `mktemp` file: `tee /dev/stderr`
+reopens fd 2, and under `./ci.sh > log 2>&1` the two offsets punch NUL holes through each other's
+lines (4d).
 
 The floors live in the `test` script, so `npm test` and the gate are one path: 100% of lines and
 functions, and a branch floor that only ever moves upward. It sits below 100 because the guards

ci.sh

@@ -3,16 +3,14 @@ set -euo pipefail
 cd "$(dirname "$0")"
 source ./docker-runner.sh
 
-test_log=$(mktemp)
+leg "install ($node_image)" in_image "$node_image" npm ci
+leg "typecheck ($node_image)" in_image "$node_image" npm run typecheck
 
-# Streamed, and copied so the zero-test guard reads the count without trading the output for it.
+test_log=$(mktemp)
 node_tests() {
   in_image "$node_image" npm test 2>&1 | tee "$test_log"
 }
 
-leg "install ($node_image)" in_image "$node_image" npm ci
-leg "typecheck ($node_image)" in_image "$node_image" npm run typecheck
-
 leg "tests ($node_image)" node_tests
 if grep -q 'ℹ tests 0' "$test_log"; then
   echo 'the gate ran zero tests — failing instead of a vacuous green'

docker-runner.sh

@@ -1,3 +1,5 @@
+: "${EPOCHREALTIME:?the gate times its legs with EPOCHREALTIME — bash 5 or newer}"
+
 bun_image=oven/bun:1.4.0-alpine
 deno_image=denoland/deno:2.9.6
 firefox_image=selenium/standalone-firefox:153.0.4
@@ -10,8 +12,7 @@ in_image() {
   docker run --rm -u "$(id -u):$(id -g)" -e HOME=/tmp ${PROPERTY_RUNS+-e PROPERTY_RUNS} ${in_image_network:+--network "$in_image_network"} -v "$PWD:/app" -w /app --entrypoint "$entrypoint" "$image" "$@"
 }
 
-# Markers go to stderr so a leg reporting a value stays capturable, and the locals carry the
-# function's own name because bash scopes them dynamically into whatever the leg runs.
+# Markers on stderr so a captured leg's value stays clean; leg_* because bash scopes local into the leg's own call.
 leg() {
   local leg_name=$1 leg_elapsed leg_started leg_status=0
   shift
@@ -26,7 +27,7 @@ leg() {
 
 with_firefox() {
   local container in_image_network status=0
-  container=$(docker run -d --rm "$firefox_image")
+  container=$(docker run -d --rm "$firefox_image") || return $?
   # The id is baked in: the trap fires after this function's locals are gone.
   trap "docker rm -f $container >/dev/null 2>&1" EXIT
   trap 'exit 130' INT

publish.sh

@@ -3,23 +3,22 @@ set -euo pipefail
 cd "$(dirname "$0")"
 source ./docker-runner.sh
 
-read_field() {
-  in_image "$node_image" npm pkg get "$1" | tr -d '"\r'
-}
-
-read_package_fields() {
-  private=$(read_field private)
-  name=$(read_field name)
-  version=$(read_field version)
-}
-
 published_version() {
   in_image "$node_image" npm view "$1@$2" version 2>/dev/null || true
 }
 
 push_tag() {
-  git tag "v$version"
-  git push origin "v$version"
+  git tag "v$version" && git push origin "v$version"
+}
+
+read_field() {
+  in_image "$node_image" npm pkg get "$1" | tr -d '"\r'
+}
+
+read_package_fields() {
+  private=$(read_field private) &&
+    name=$(read_field name) &&
+    version=$(read_field version)
 }
 
 leg "read package.json ($node_image)" read_package_fields
@@ -28,7 +27,7 @@ if [ "$private" = 'true' ]; then
   exit 0
 fi
 
-published=$(leg "ask npmjs for $name@$version" published_version "$name" "$version")
+published=$(leg "ask npmjs for $name@$version ($node_image)" published_version "$name" "$version")
 tagged=$(leg "ask origin for v$version" git ls-remote --tags origin "v$version")
 
 # Both steps observe their own end state, so a partial run converges on the next push to main.

todo-history.md

@@ -560,19 +560,27 @@ The done `todo.md` items in full, as they were written. `todo.md` keeps a one-li
       rather than promoting the gate's `dist`, and unmeasured until the log shows them. Per-leg
       timing is what turns "slow or hung" from a guess into a reading; the browser leg's own
       5.4–7.9s against a 17s warm gate is the number that made it obviously cheap.
-      **Measured** (2026-09-19): ten legs, not the nine counted above, each naming its image, on a
-      27.9 s warm gate — install 1.5 s, typecheck 1.1 s, Node tests 4.6 s, Deno 6.3 s, Bun 3.9 s,
-      build 1.0 s, pack and install 1.6 s, consumer typecheck 1.0 s, engines floor 0.5 s, browser
-      5.7 s. The browser leg lands in the 5.4–7.9 s the item quotes, and the markers cost nothing
-      measurable: 28.9 s before against 27.9 s after. `publish.sh` reads its fields in 2.5 s and the
-      registry in 1.4 s; its `npm ci` and rebuild are the gate's own 1.5 s and 1.0 s, so the seconds
-      §9 accepts for rebuilding rather than promoting the gate's `dist` are about 2.5.
-      Three things the writing turned up. The markers print to stderr, so a leg whose value is read —
-      `publish.sh` asking npmjs — stays capturable. `leg`'s locals carry its own name because bash
-      scopes them into whatever the leg runs: unprefixed, `name` was swallowed by the leg reading
-      `package.json`. And `leg` returns its command's status the way `with_firefox` already did,
-      because the bare call swallowed a non-zero one wherever `set -e` is suspended, which also gets
-      the elapsed time printed for the leg that failed.
+      **Measured** (2026-09-20): ten legs, not the nine counted above, each naming the image it runs
+      in where it runs in one, on a 28.4 s warm gate — install 1.5 s, typecheck 1.2 s, Node tests
+      4.9 s, Deno 6.0 s, Bun 4.5 s, build 1.0 s, pack and install 1.6 s, consumer typecheck 1.0 s,
+      engines floor 0.4 s, browser 5.6 s. The browser leg lands in the 5.4–7.9 s the item quotes,
+      and the markers cost nothing measurable: 28.9 s before against 28.4 s after. `publish.sh`
+      reads its fields in 2.6 s and the registry in 1.4 s; its `npm ci` and rebuild are the gate's
+      own 1.5 s and 1.0 s, so the seconds §9 accepts for rebuilding rather than promoting the gate's
+      `dist` are about 2.5.
+      Four things the writing turned up, three of them bash scoping a rule differently than it
+      reads. The markers print to stderr, so a leg whose value is read — `publish.sh` asking npmjs —
+      stays capturable. `leg`'s locals carry its own name because bash scopes them into whatever the
+      leg runs: unprefixed, `name` was swallowed by the leg reading `package.json`. `leg` returns
+      its command's status the way `with_firefox` already did, because the bare call dropped a
+      non-zero one wherever `set -e` is suspended, which also gets the elapsed time printed for the
+      leg that failed. And the `||` that captures that status suspends `set -e` for everything the
+      leg calls, so a function a leg runs chains its statements with `&&` or every statement but the
+      last runs unchecked: `read_package_fields` read on past a failed read, and `push_tag` pushed a
+      tag the tag step had refused to write, both of which aborted before this chunk (the
+      stability-reviewer, 2026-09-20). §10 carries the rule so the next leg cannot reintroduce it,
+      and `EPOCHREALTIME` is guarded at `source` so an older bash names itself rather than dying as
+      an unbound variable on the first leg.
 
 - [x] **5a — Rename to `@larvit/adf-codec` (`0.1.0`).** Before the first publish, the name being
       the published identity: `package.json` `name` and `repository`, the Gitea repo and its