esp32s3: fresh build of FastLED Blink now selects Arduino's Matter library (image 462 KB → 1.6 MB), same commit and fbuild 2.5.24

[zackees]

Downstream: FastLED/FastLED#4474 (the ESP32-S3 bloat gate is red on master).

Repro

FastLED's bloat_regression_esp32s3.yml runs bash bloat esp32s3 --build on a clean runner (no fbuild cache restore), with fbuild==2.5.24.

• Run 35436746934 attempt 1 (2026-09-19 10:12Z), commit 8999762397: framework core cache: miss key=07cf2eef0eca, 59 core files, image 462.49 KB.
• The same run re-run as attempt 2 (~17:20Z), same commit and same fbuild: framework core cache: miss key=85bbccbc1fa1, 60 core files, image 1.60 MB.

The extra flash is Arduino's Matter library (libmatter.a) and everything it references: libespressif__esp_matter.a (+334 KB), net80211, lwip, bt/btdm, mbedcrypto, wpa_supplicant, mdns. Blink uses none of them. So library selection (lib_ldf_mode = chain) now resolves the Matter library that it didn't resolve seven hours earlier.

Inputs

fbuild install -e esp32s3 --dry-run resolves pinned packages: platform 54.03.20, esp32-arduino 3.3.11 core and libs, toolchain xtensa-esp-elf-14.2.0_20260121. It also lists a platform entry from the floating https://github.com/pioarduino/platform-espressif32/releases/download/stable/platform-espressif32.zip, the only unpinned input visible. If library selection, or the core key, reads anything from the stable platform, a new stable release would explain a flip on an unchanged commit.

A build with a warm local cache (keys from before the flip) still produces 473,524 B.

Ask

• Where does the core key 85bbccbc1fa1 differ from 07cf2eef0eca, and why does the library scan now select Matter?
• Pin, or stop consulting, the floating stable platform if that's the cause.

🤖 Generated with Claude Code

[zackees]

It doesn't reproduce locally from an empty cache. An isolated daemon (FBUILD_CACHE_DIR=<empty dir>, FBUILD_DAEMON_PORT=60611, fbuild 2.5.24, NixOS) built FastLED master Blink for esp32s3 at 2026-09-19 ~18:00Z. It downloaded everything fresh and compiled 60 core files, the same count as the failing CI runs, but the image is 473,644 B with no Matter. So the 60-file core isn't the cause. The Matter library selection happens only on the GitHub ubuntu-latest runner, and it started between 10:12Z and ~16:20Z on an unchanged commit.

[zackees]

Ruled out so far:

• Transient: a third CI attempt (run 35455732120, ~18:20Z) fails identically, with core key 85bbccbc1fa1 and 1.60 MB.
• fbuild self-update: no release after v2.5.24 (2026-09-18 19:36Z), and CI installs 2.5.24.
• Runner image: passing and failing attempts report the same image versions (20260828.587 / 20260907.300.1).
• FastLED/boards main: no commits since 08:00Z.
• Fresh local cache: doesn't reproduce (473,644 B).

Next useful datum: fbuild's library-selection result on the runner, i.e. which Arduino libraries it resolved and why Matter was picked. It doesn't appear in the default build output. Is there a flag or env var that logs it?