[thinkyhead]

Description

"PR Test Build" only builds the folders a PR changes. When Marlin itself grows, an example that no PR touches can stop fitting in flash, and nobody sees it until the next PR to touch that example fails. Nothing records how large each example is either.

This adds a scheduled workflow (daily 07:17 UTC, or by hand) that builds every example against the current bugfix-2.1.x:

• plan: splits the 422 folders into shards of at most 30 (24 shards today). A shard holds one toolchain group, so it uses the cache that cache-pio.yml already saves.
• build: each shard runs build-configs. All shards use the same Marlin commit, even if bugfix-2.1.x moves during the run.
• manifest: merges the results into sizes.tsv (the sizes artifact, 90 days), compares it with the previous run, and writes the job summary.

The summary lists failures (marking new ones), configs fixed or added since the previous run, the median size change per environment (Marlin's own growth, not one example's), the biggest changes in a single config, the configs within 2 KB of the flash limit, and a link to the Marlin commits since the last run.

The run fails if a config fails to build or a shard doesn't finish. It needs no write permission, only actions: read to find the previous manifest. A run started by hand takes a regex to build only some folders (saved as sizes-partial, so it isn't used as the baseline).

The planning, merging and report are in .github/scripts/size-manifest.

Testing

• Planned against the real 422 folders: 24 shards, 12 toolchain groups.
• Ran on a fork twice with a four-pattern filter (7 configs, 3 shards): the first run built all 7, the second compared with it and reported no change. My first trial found a bug (every shard saved results.tsv, so merging overwrote them), which is fixed.
• actionlint is clean.
• Not yet run here: the full build of all 422 configs, or the scheduled trigger. Expect the first full run to take a few hours of runner time across the shards.