fix: strip the shipped payload and mirror assets concurrently (2026.7.29.1)

The GitCode mirror leg of `publish-ecosystem` has failed four releases in a
row — 0.0.94 / 0.0.97 / 0.0.105 / 2026.7.28.2 — each time on the biggest
tarballs, each time costing ~5 minutes of manual `gtc release upload`. The
per-asset 180s cap was tuned three times without anyone measuring what it was
capping. Probe PR #301 measured it:

  GitHub US runner -> file.gitcode.com   0.012 MB/s   <- the failing path
  same runner      <- file.gitcode.com   3.87  MB/s
  same runner      -> github.com        16     MB/s
  mainland-CN host -> file.gitcode.com   1.84  MB/s

file.gitcode.com is a single Huawei Cloud origin in Beijing, so it is the
inbound-to-CN direction that is shaped — not the host (the same runner
downloads from it at 3.87 MB/s), not the client (curl and urllib measure
identically), and not the runner's egress (16 MB/s to GitHub). The rate also
swings ~4.6x run to run, so no fixed per-asset cap can be both safe and
useful. At 0.012 MB/s a 34.8MB asset needs ~45 minutes; 180s never had a
chance.

Two fixes, both measured rather than guessed.

S1 — stop shipping 30MB of debug info. Of the 34.8MB linux-x86_64 tarball,
almost none was mcpp: the vendored `registry/bin/xlings` shipped at 97.3MB
with full debug info (86.9MB on aarch64) against a 4.3MB mcpp. Stripping both
takes the tarball to 4.62MB — 7.5x — and the stripped binaries were verified
working (`--version`, `new`, `build`, `run`). Two defects made this possible:

  * the vendored xlings was never stripped at any of the four injection sites,
    only `cp`'d;
  * the x86_64 `strip` that DID exist ran before `mcpp pack`, which rebuilds
    the binary and overwrote it — which is why every release up to and
    including 2026.7.28.2 shipped an unstripped bin/mcpp while aarch64 (which
    stages by hand) shipped a stripped one.

tools/slim_linux_payload.sh now strips both binaries on the STAGED payload and
ASSERTS the result, so a strip that silently stops working fails the release
instead of quietly shipping a fat tarball again. The aarch64 leg's
best-effort `|| true` strip is now a hard requirement. macOS and Windows are
deliberately left alone: their payloads are already 6.1MB / 4.2MB, and
stripping a Mach-O invalidates its ad-hoc signature.

S2 — mirror a host's assets concurrently, and bound the LEG instead of each
asset. The shaping is per-connection: 1/4/8 concurrent 1MB uploads took
76s/80s/93s wall (6.6x aggregate at N=8). Assets within a leg now upload in
parallel under MIRROR_MAX_PARALLEL, and MIRROR_UPLOAD_TIMEOUT is replaced by
MIRROR_LEG_DEADLINE_GH/GTC. The v0.0.90 rule still holds — nothing is killed
and retried inside a round, because the presigned OBS PUT has no
multipart/resume (confirmed in #301); exhausting the budget ends the leg and
the completeness gate fails loudly as before.

Together the gitcode leg goes from ~45 min plus a manual step to ~6 min
unattended, and every user's download shrinks 7.5x.

Verified locally: slim on the real 2026.7.28.2 payload (101.6MB -> 13.9MB of
binaries, 34.81MB -> 4.62MB tarball, stripped mcpp builds and runs an
`import std` project); an isolated harness for the concurrent launcher
(bounded concurrency, ordered log replay, correct per-asset success/failure);
a full real mirror_res.sh run against the existing 2026.7.28.2 tag (all
skipped, gate 16/16); and a real parallel-upload run on a throwaway
`0.0.0-mirrortest` tag against BOTH hosts (gate 8/8, exit 0), whose artifacts
were deleted afterwards with the real release verified intact.

NB: release.yml is not exercised by PR CI — it runs on tag/dispatch only — so
S1 lands its first real proof at the next release, where the new assertions
fail loudly rather than silently.

Author: Sunrisepeak

.github/tools/mirror_res.sh

@@ -42,24 +42,39 @@ read -r -a ASSETS <<< "${ASSETS:-$DEFAULT_ASSETS}"
 
 info() { echo "[mirror] $*"; }
 
-# Per-asset upload cap. Exceeding it WARNs and abandons that asset (no retry,
-# no delete) so one slow cross-border PUT can't eat the whole job budget — the
-# v0.0.94 release lost `publish-ecosystem` to a 30min job timeout with zero
-# per-asset visibility, and the two biggest linux tarballs had to be pushed by
-# hand afterwards.
+# ── Budget: one deadline per HOST LEG, not per asset ────────────────────────
+# The old per-asset cap (180s, abandon-on-expiry) failed four releases in a row
+# — 0.0.94 / 0.0.97 / 0.0.105 / 2026.7.28.2 — and was tuned three times without
+# anyone measuring what it was capping. Measured (probe PR #301):
 #
-# CAUTION (v0.0.90 postmortem, still binding): a cap that kills a
-# slow-but-PROGRESSING upload and then RETRIES it is strictly worse than no cap
-# — every restart resumes from byte zero and the mirror never converges. This
-# cap is safe only because a capped asset is SKIPPED, never re-uploaded. The
-# completeness gate at the bottom is still the pass/fail, so a skipped asset
-# fails the release loudly instead of silently shipping a half mirror.
+#   GitHub US runner -> file.gitcode.com  0.012 MB/s   <- the failing path
+#   same runner      <- file.gitcode.com  3.87  MB/s
+#   same runner      -> github.com       16     MB/s
+#   mainland-CN host -> file.gitcode.com  1.84  MB/s
 #
-# Sizing: mcpp's largest asset is ~30MB (no package exceeds 100MB). 180s is a
-# generous ceiling for that — an upload still running at 3min is not "slow", it
-# is stuck, and the right move is to stop paying CI for it and push that one
-# asset by hand (the gate below prints the exact command).
-: "${MIRROR_UPLOAD_TIMEOUT:=180}"
+# file.gitcode.com is a single Huawei Cloud origin in Beijing; it is the
+# inbound-to-CN direction that is shaped, not the host and not the client
+# (curl and urllib measure identically). At 0.012 MB/s a 34.8MB asset needs
+# ~45 MINUTES, so no per-asset value in the 180s neighbourhood was ever going
+# to work. The rate also varies ~4.6x run to run (0.011-0.051 MB/s), which
+# rules out ANY fixed per-asset cap being simultaneously safe and useful.
+#
+# So: bound the LEG, and let each upload have whatever is left of it. A leg
+# that overruns fails the completeness gate exactly as before.
+#
+# The v0.0.90 rule still binds: never kill a slow-but-PROGRESSING upload and
+# then retry it — a restart resumes from byte zero (the presigned OBS PUT has
+# no multipart/resume; see #301). Here nothing is killed and retried within a
+# round: the deadline ends the leg.
+: "${MIRROR_LEG_DEADLINE_GH:=600}"     # github is fast; 10min is already absurd
+: "${MIRROR_LEG_DEADLINE_GTC:=2400}"   # gitcode: shaped inbound, needs headroom
+
+# Assets are uploaded CONCURRENTLY within a leg. Measured on the same probe:
+# the shaping is per-CONNECTION, so concurrency scales almost linearly —
+# 1/4/8 concurrent 1MB uploads took 76s/80s/93s wall (0.013/0.050/0.086 MB/s
+# aggregate, 6.6x at N=8). Concurrency does raise the error rate (one 502 in
+# the N=4 round), which the existing probe-then-reupload rounds absorb.
+: "${MIRROR_MAX_PARALLEL:=8}"
 
 DL="$(mktemp -d)"; trap 'rm -rf "$DL"' EXIT
 
@@ -70,24 +85,34 @@ human_size() { # path → e.g. 31.9MB
 
 host_label() { [[ "$1" == gh ]] && echo github || echo gitcode; }
 
-# Upload one asset under the cap, timing it. 0 = the command returned within
-# the cap (NOT proof it landed — gtc's exit code lies both ways, so the probe /
-# gate remains the only source of truth); 1 = the cap fired, asset abandoned.
-upload_asset() { # kind(gh|gtc) asset → 0 returned / 1 capped
-  local kind="$1" a="$2" start elapsed rc=0 sz host
+# Upload one asset with whatever is left of the leg deadline, timing it.
+# 0 = the command returned inside the budget (NOT proof it landed — gtc's exit
+# code lies both ways, so the probe / gate remains the only source of truth);
+# 1 = the leg deadline fired.
+#
+# Runs in a SUBSHELL under the parallel launcher, so it must not rely on any
+# state surviving the call; everything it produces goes to stdout/stderr, which
+# the launcher collects per asset and replays in order.
+upload_asset() { # kind(gh|gtc) asset deadline_epoch → 0 returned / 1 out of budget
+  local kind="$1" a="$2" deadline="$3" start elapsed rc=0 sz host budget
   sz=$(human_size "$DL/$a")
   host=$(host_label "$kind")
+  budget=$(( deadline - SECONDS ))
+  if (( budget <= 0 )); then
+    info "WARN: $host $a ($sz) not attempted — leg deadline already spent"
+    return 1
+  fi
   start=$SECONDS
   if [[ "$kind" == gh ]]; then
-    GH_TOKEN="${XLINGS_RES_TOKEN:-}" timeout "$MIRROR_UPLOAD_TIMEOUT" \
+    GH_TOKEN="${XLINGS_RES_TOKEN:-}" timeout "$budget" \
       gh release upload "$VER" "$DL/$a" -R "$GH_DST" --clobber >/dev/null 2>&1 || rc=$?
   else
-    timeout "$MIRROR_UPLOAD_TIMEOUT" gtc release upload "$GTC_DST" "$DL/$a" --tag "$VER" \
+    timeout "$budget" gtc release upload "$GTC_DST" "$DL/$a" --tag "$VER" \
       >/dev/null 2>&1 || rc=$?
   fi
   elapsed=$((SECONDS - start))
   if [[ $rc == 124 || $rc == 137 ]]; then
-    info "WARN: $host $a ($sz) exceeded the ${MIRROR_UPLOAD_TIMEOUT}s cap after ${elapsed}s — skipping (not retried; the verify gate below decides the release)"
+    info "WARN: $host $a ($sz) hit the leg deadline after ${elapsed}s — abandoning (the verify gate below decides the release)"
     return 1
   fi
   info "$host $a ($sz) uploaded in ${elapsed}s"
@@ -146,39 +171,73 @@ verify_batch() { # base_url asset... → prints assets still not serving
 #  - NEVER kill a slow-but-progressing upload AND RETRY IT. v0.0.90 wrapped
 #    uploads in `timeout 300`, so every cross-border PUT >5min was SIGKILLed
 #    at 60%% and restarted from byte zero — the 20min job ceiling fell to
-#    this. MIRROR_UPLOAD_TIMEOUT keeps a cap but ABANDONS the asset instead of
-#    retrying it, which is what makes the cap safe; the gate then fails the
-#    release loudly rather than thrashing until the job is killed.
+#    this. Nothing is killed-and-retried inside a round now: an upload gets
+#    the remaining LEG budget, and exhausting it ends the leg.
 #  - gtc's exit code lies both ways (obs_callback flakiness); the download
-#    probe is the only source of truth.
-mirror_host() { # kind(gh|gtc) base_url
-  local kind="$1" base="$2" try a
+#    probe is the only source of truth. Measured mechanism (#301): the
+#    presigned PUT carries an `x-obs-callback` header pointing at
+#    api.gitcode.com; OBS stores the object and THEN calls back, so a failed
+#    callback reports `code:400 ... EOF` for an upload that did land.
+#
+# Assets within a leg upload CONCURRENTLY (the shaping is per-connection —
+# see the deadline block at the top). Each upload runs in its own subshell
+# writing to its own log, which is replayed in asset order after the wait, so
+# concurrent progress lines don't interleave into unreadable soup.
+mirror_host() { # kind(gh|gtc) base_url deadline_seconds
+  local kind="$1" base="$2" budget="$3" try a
   local pending failed
-  local -A capped=()
+  local -A lost=()
   local host_start=$SECONDS
+  local deadline=$((SECONDS + budget))
   local host; host=$(host_label "$kind")
+  local wdir; wdir=$(mktemp -d)
+  info "$host leg budget ${budget}s, up to ${MIRROR_MAX_PARALLEL} concurrent uploads"
   for try in 1 2 3; do
-    pending=()
+    # Step 1: probe first — anything already serving is mirrored and must
+    # never be re-uploaded.
+    local todo=()
     for a in "${ASSETS[@]}"; do
-      # A capped asset is never re-attempted (see MIRROR_UPLOAD_TIMEOUT).
-      [[ -n "${capped[$a]:-}" ]] && continue
-      # Step 1: already serving? then it's mirrored — never re-upload it.
+      [[ -n "${lost[$a]:-}" ]] && continue
       if probe "${base}/${a}"; then
         [[ $try == 1 ]] && info "$host $a already mirrored, skipping"
         continue
       fi
-      if upload_asset "$kind" "$a"; then
+      todo+=("$a")
+    done
+    [[ ${#todo[@]} == 0 ]] && break
+
+    # Step 2: upload the missing ones concurrently, bounded by a live count of
+    # running children (not a hand-kept counter — that mis-books as soon as one
+    # finishes early). `if/else` around upload_asset so the subshell's own
+    # errexit can't swallow the rc file.
+    local i=0
+    for a in "${todo[@]}"; do
+      while (( $(jobs -rp | wc -l) >= MIRROR_MAX_PARALLEL )); do sleep 1; done
+      ( if upload_asset "$kind" "$a" "$deadline" >"$wdir/$i.log" 2>&1
+        then echo 0; else echo 1; fi >"$wdir/$i.rc" ) &
+      i=$((i + 1))
+    done
+    wait
+
+    pending=()
+    for i in "${!todo[@]}"; do
+      a="${todo[$i]}"
+      cat "$wdir/$i.log" 2>/dev/null || true
+      if [[ "$(cat "$wdir/$i.rc" 2>/dev/null || echo 1)" == 0 ]]; then
         pending+=("$a")
       else
-        capped[$a]=1
+        lost[$a]=1
       fi
+      rm -f "$wdir/$i.log" "$wdir/$i.rc"
     done
+
     [[ ${#pending[@]} == 0 ]] && break
     failed=$(verify_batch "$base" "${pending[@]}")
     [[ -z "$failed" ]] && break
     info "$host not serving after patience (try $try): $failed — re-uploading (no delete)"
   done
-  ((${#capped[@]})) && info "WARN: $host abandoned ${#capped[@]} asset(s) at the ${MIRROR_UPLOAD_TIMEOUT}s cap: ${!capped[*]}"
+  rm -rf "$wdir"
+  ((${#lost[@]})) && info "WARN: $host abandoned ${#lost[@]} asset(s) at the ${budget}s leg deadline: ${!lost[*]}"
   info "$host mirror leg finished in $((SECONDS - host_start))s"
   return 0  # the completeness gate below is the real pass/fail
 }
@@ -205,11 +264,11 @@ fi
 
 GH_PID=""; GTC_PID=""
 if [[ "$GH_ENABLED" == 1 ]]; then
-  mirror_host gh  "https://github.com/${GH_DST}/releases/download/${VER}" &
+  mirror_host gh  "https://github.com/${GH_DST}/releases/download/${VER}" "$MIRROR_LEG_DEADLINE_GH" &
   GH_PID=$!
 fi
 if [[ "$GTC_ENABLED" == 1 ]]; then
-  mirror_host gtc "https://gitcode.com/${GTC_DST}/releases/download/${VER}" &
+  mirror_host gtc "https://gitcode.com/${GTC_DST}/releases/download/${VER}" "$MIRROR_LEG_DEADLINE_GTC" &
   GTC_PID=$!
 fi
 [[ -n "$GH_PID"  ]] && wait "$GH_PID"
@@ -232,7 +291,7 @@ for host in "${hosts[@]}"; do
   done
 done
 if [[ $rc != 0 ]]; then
-  echo "[mirror] hint: if the asset above was WARNed as capped at ${MIRROR_UPLOAD_TIMEOUT}s, either raise MIRROR_UPLOAD_TIMEOUT for this run or push it by hand:" >&2
+  echo "[mirror] hint: if the asset above was WARNed at a leg deadline, raise MIRROR_LEG_DEADLINE_GH/GTC for this run, or push it by hand:" >&2
   echo "[mirror]       gh release download v$VER -R $SRC_REPO -p '<asset>' && gtc release upload $GTC_DST '<asset>' --tag $VER" >&2
 fi
 [[ $rc == 0 ]] && info "all assets mirrored + verified on ${#hosts[@]} host(s) in ${SECONDS}s"

.github/tools/slim_linux_payload.sh

@@ -0,0 +1,70 @@
+#!/usr/bin/env bash
+# slim_linux_payload.sh — strip the shipped ELF binaries in a staged release
+# wrapper directory, then ASSERT the result.
+#
+# WHY. The linux tarballs were 34.8MB (x86_64) and 30.2MB (aarch64) while the
+# macOS one was 6.1MB and the Windows one 4.2MB. Almost none of that was mcpp:
+# the vendored `registry/bin/xlings` shipped at 97.3MB / 86.9MB with full debug
+# info, against a 4.3MB / 2.8MB mcpp. Stripping both takes the x86_64 tarball
+# from 34.81MB to 4.62MB — 7.5x — with the stripped binaries verified working
+# (`--version`, `new`, `build`, `run`).
+#
+# That size is what makes the GitCode mirror leg fail: measured upload from a
+# GitHub-hosted US runner to file.gitcode.com (a single Huawei Cloud origin in
+# Beijing) is ~0.012 MB/s in the inbound-to-CN direction — the SAME runner
+# downloads from that host at 3.87 MB/s and uploads to GitHub at 16 MB/s, and a
+# mainland-CN host reaches 1.84 MB/s to the same endpoint. At 0.012 MB/s a
+# 34.8MB asset needs ~45 minutes; at 4.6MB it needs ~6.
+#
+# ASSERT, don't trust. release.yml already stripped the x86_64 mcpp before
+# packing — and shipped it unstripped anyway, because `mcpp pack` rebuilds the
+# binary and overwrites the stripped one. A `strip` whose effect is never
+# checked is a comment, not a step. Hence: this runs on the STAGED payload
+# (after pack, before tar) and fails loudly if anything is still unstripped.
+#
+# Linux only, on purpose. The macOS and Windows payloads are already small, and
+# stripping a Mach-O invalidates its (ad-hoc) code signature — not worth the
+# risk for ~2MB when 100% of the mirror problem is the two linux tarballs.
+#
+# Usage: slim_linux_payload.sh <wrapper-dir> [strip-cmd]
+set -euo pipefail
+
+DIR="${1:?usage: slim_linux_payload.sh <wrapper-dir> [strip-cmd]}"
+STRIP="${2:-strip}"
+
+[ -d "$DIR" ] || { echo "slim: no such directory: $DIR" >&2; exit 1; }
+command -v "$STRIP" >/dev/null 2>&1 || [ -x "$STRIP" ] || {
+  echo "slim: strip tool not usable: $STRIP" >&2; exit 1; }
+
+size_of() { stat -c %s "$1" 2>/dev/null || stat -f %z "$1"; }
+mb() { awk -v b="$1" 'BEGIN{printf "%.1f", b/1048576}'; }
+
+# The two binaries the release actually ships. Both are static ELFs; a missing
+# one is not an error (the aarch64 leg skips xlings when the upstream download
+# fails, and that has its own handling), but a PRESENT one must end up stripped.
+total_before=0 total_after=0 found=0
+for rel in bin/mcpp registry/bin/xlings; do
+  f="$DIR/$rel"
+  [ -f "$f" ] || { echo "slim: $rel absent, skipping"; continue; }
+  found=$((found + 1))
+  before=$(size_of "$f")
+  "$STRIP" --strip-unneeded "$f" 2>/dev/null || "$STRIP" "$f"
+  after=$(size_of "$f")
+  total_before=$((total_before + before))
+  total_after=$((total_after + after))
+  echo "slim: $rel  $(mb "$before")MB -> $(mb "$after")MB"
+
+  # The assertion this script exists for.
+  if file "$f" | grep -q 'not stripped'; then
+    file "$f"
+    echo "slim: FAIL: $rel is still not stripped after running $STRIP" >&2
+    exit 1
+  fi
+  # A stripped binary that no longer runs is worse than a fat one. The callers
+  # smoke-test the packaged tarball too, but catch it here where the failure
+  # names the file.
+  [ -x "$f" ] || { echo "slim: FAIL: $rel lost its exec bit" >&2; exit 1; }
+done
+
+[ "$found" -gt 0 ] || { echo "slim: FAIL: no shippable binary found under $DIR" >&2; exit 1; }
+echo "slim: payload $(mb "$total_before")MB -> $(mb "$total_after")MB across $found binary(ies)"

.github/workflows/release.yml

@@ -132,8 +132,11 @@ jobs:
           ARTIFACT=$(find target/x86_64-linux-musl -type f -name mcpp | head -1)
           test -n "$ARTIFACT"
           file "$ARTIFACT" | grep -q 'statically linked'
-          # Strip — debug info on a static ELF balloons it ~7×.
-          strip "$ARTIFACT"
+          # NB: stripping here is pointless — `mcpp pack` below rebuilds the
+          # binary and overwrites it, which is why every release up to and
+          # including 2026.7.28.2 shipped an UNSTRIPPED bin/mcpp despite the
+          # strip that used to live on this line. The payload is slimmed after
+          # packing instead, in the inject step below.
 
           # Pack with the freshly-built mcpp (not the bootstrap) so any
           # fixes to the pack code path are exercised in the same release
@@ -156,6 +159,10 @@ jobs:
           mkdir -p "$INJECT/$WRAPPER/registry/bin"
           cp "$XLINGS_BIN" "$INJECT/$WRAPPER/registry/bin/xlings"
           chmod +x "$INJECT/$WRAPPER/registry/bin/xlings"
+          # Slim AFTER pack (pack rebuilds bin/mcpp) and BEFORE tar. Asserts
+          # the result, so a strip that silently stops working fails the
+          # release instead of quietly shipping a 34.8MB tarball again.
+          bash .github/tools/slim_linux_payload.sh "$INJECT/$WRAPPER"
           (cd "$INJECT" && tar -czf "$GITHUB_WORKSPACE/${TARBALL}" "$WRAPPER")
           rm -rf "$INJECT"
 
@@ -330,9 +337,15 @@ jobs:
           STAGING=$(mktemp -d)
           mkdir -p "$STAGING/$WRAPPER/bin"
           cp "$MCPP_AARCH64" "$STAGING/$WRAPPER/bin/mcpp"
-          # Strip with the cross toolchain's strip if present (binary is aarch64).
+          # The binary is aarch64, so the host x86_64 `strip` cannot touch it —
+          # resolve a cross-capable one. Required, not best-effort: the old
+          # `|| true` here meant a missing tool silently shipped a fat tarball.
+          # (slim_linux_payload.sh runs after xlings is staged, below.)
           STRIP=$(find "$HOME/.mcpp" -name 'aarch64-linux-musl-strip' -type f 2>/dev/null | head -1)
-          [ -n "$STRIP" ] && "$STRIP" "$STAGING/$WRAPPER/bin/mcpp" 2>/dev/null || true
+          [ -n "$STRIP" ] || STRIP=$(command -v llvm-strip 2>/dev/null || true)
+          [ -n "$STRIP" ] || STRIP=$(command -v aarch64-linux-gnu-strip 2>/dev/null || true)
+          [ -n "$STRIP" ] || { echo "no aarch64-capable strip found"; exit 1; }
+          echo "aarch64 strip: $STRIP"
           cp LICENSE "$STAGING/$WRAPPER/" 2>/dev/null || true
           cp README.md "$STAGING/$WRAPPER/" 2>/dev/null || true
           cat > "$STAGING/$WRAPPER/mcpp" << 'LAUNCHER'
@@ -356,6 +369,9 @@ jobs:
               chmod +x "$STAGING/$WRAPPER/registry/bin/xlings"
             fi
           fi
+          # Slim both shipped ELFs with the cross strip resolved above, and
+          # assert the result (the vendored xlings was 86.9MB unstripped here).
+          bash .github/tools/slim_linux_payload.sh "$STAGING/$WRAPPER" "$STRIP"
           mkdir -p dist
           (cd "$STAGING" && tar -czf "$GITHUB_WORKSPACE/dist/${TARBALL_NAME}" "$WRAPPER")
           cp "dist/${TARBALL_NAME}" "dist/mcpp-linux-aarch64.tar.gz"

mcpp.toml

@@ -1,6 +1,6 @@
 [package]
 name        = "mcpp"
-version     = "2026.7.28.2"
+version     = "2026.7.29.1"
 description = "Modern C++ build & package management tool"
 license     = "Apache-2.0"
 authors     = ["mcpp-community"]

src/toolchain/fingerprint.cppm

@@ -18,7 +18,7 @@ import mcpp.toolchain.detect;
 
 export namespace mcpp::toolchain {
 
-inline constexpr std::string_view MCPP_VERSION = "2026.7.28.2";
+inline constexpr std::string_view MCPP_VERSION = "2026.7.29.1";
 
 struct FingerprintInputs {
     Toolchain                       toolchain;