fix(ci): report flash and static RAM usage separately

[dakejahl]

Summary

Report separate flash and static RAM deltas while retaining Bloaty's section and symbol breakdown.

Problem

The current headline treats total VM growth as flash growth. For #28760, FMUv5x shows +7,608 B even though flash grows by +3,128 B and static RAM by +4,480 B.

Solution

Measure the flash load image and RAM allocations separately from ELF section headers, with no per-target constants: sections with VMA == LMA execute from flash, and everything else allocated counts as RAM. Initialized data and RAM code count in both totals, flash includes padding exactly as objcopy -O binary does, and debug information is excluded. Program headers aren't used because, depending on ld's page size, the first LOAD segment can include the ELF header below the flash origin. Growth over 100 B is marked yellow and over 1000 B red; savings over 100 B are marked green. The Bloaty breakdown is limited to VM sizes, which drops the debug and symbol table rows. A new analysis comment is posted only when flash or static RAM changes on a target. Existing comments are updated even when every delta returns to zero.

PRs are diffed against the first parent of the merge commit instead of the base branch tip. If the PR merged while the job was queued, the tip already contains it, which caused the near-zero diffs seen after merge.

Validated with eight ARM ELF tests, flash totals matching the .bin for ark_fmu-v6x_default and px4_fmu-v6x_flash-analysis, and the CI bloaty-action image with --domain=vm.

Comment preview: #28760 linker totals; breakdown from a local fmu-v6x change adding a 4 KiB buffer

🔎 Flash and RAM Analysis

Target	Flash change	Static RAM change
px4_fmu-v5x	🔴 +3,128 B (+0.15%)	🔴 +4,480 B (+3.82%)
px4_fmu-v6x	🔴 +3,120 B (+0.16%)	🔴 +4,480 B (+3.95%)

px4_fmu-v6x — Bloaty section/symbol breakdown

     VM SIZE    
 -------------- 
  +4.0% +4.00Ki    .bss
    [NEW] +4.00Ki    probe_bss
  +0.0%    +360    .text
    [NEW]    +256    probe_data
     +11%    +108    ver_main
     +44%      +4    g_nullstring
   +10.0%      +2    DEFAULT_NETMAN_CONFIG
   -33.3%      -2    sz_ver_hw_str
    -0.0%      -4    [section .text]
    -0.5%      -4    up_dshot_init
  +0.2% +4.35Ki    TOTAL

[JonasPerolini]

Thanks for this PR @dakejahl!

Claude review on behalf of @JonasPerolini

Tools/ci/firmware_size.py:29 — the flash total is derived from PT_LOAD segments, which makes it depend on ld's page-size heuristic. When the first section leaves room in its page, ld puts the ELF header into the first LOAD segment, so the segment LMA starts below the flash origin. With binutils 2.34 (GCC 9 2020-q2) and the fmu-v5x flash-analysis scripts (origin 0x08008000, 64 KB max page size) the first LOAD is at 0x08000000, the region check raises "load image outside the configured flash region", and 5 of 7 tests fail. It passes on 2.42 (px4-dev container) only because the default max page size there is 4 KB; relinking on 2.34 with -z max-page-size=0x1000 makes the failure disappear. Suggest measuring like objcopy does (ALLOC+LOAD sections via objdump -h, or just stat the .bin the build already writes) or clamping the segment start at the flash origin, plus a test that links the fixture with -z max-page-size=0x10000 to cover this layout.

Tools/ci/test_firmware_size.py:117 — the CLI test runs python3 from PATH rather than sys.executable. The script needs 3.9+ for dict[str, int], so this test fails on machines whose default python3 is older while the rest of the suite passes under a newer interpreter.

[mbjd]

Thanks! While we're at it:

• Often after a PR is merged the flash diff is updated to basically zero, such as here. Any idea why that is?
• WDYT about making it yellow if >100B, red if >1000B, green if <-100B or similar?
• We could also suggest an AI review with a prompt focused on identifying added flash bloat, or provide a link to directly provide that review, e.g. with https://claude.ai/new?q=<prompt> et al.

[dakejahl]

We could also suggest an AI review with a prompt focused on identifying added flash bloat, or provide a link to directly provide that review, e.g. with https://claude.ai/new?q= et al.

This would be cool. Deferring until we setup the infrastructure for automated AI reviews.

[julianoes]

Nice!