[mottihoresh]

Problem

library_dependents in .github/firmware-platformio-impact.json is hand-written and had drifted to roughly a fifth of the real dependency graph:

• it listed a Firmware/Prototype/ tree that does not exist and has no history in this repo;
• it had no NodeStatus key at all;
• it listed 2 of A4EC's 9 dependents.

The first two fail loud — a missing project and an unknown library both hit the selector's escape hatch, so a CANProtocol or NodeStatus change fell back to a full build. A4EC's failure is silent: the selector returned 2 projects and skipped the other 7.

Why hand-maintenance can't fix it

The graph is not in the platformio.ini files alone. Each Firmware/Libraries/*/library.json declares its own dependencies:

CANProtocol -> HIDControls, NodeStatus, STM32Board
PanelBridge -> A4EC, CANProtocol, NodeStatus, STM32Board
PanelGroup  -> CANProtocol, STM32Board
DrumDisplay -> NodeStatus, PanelGroup

A project listing only PanelGroup still compiles four libraries. Reading lib_deps alone under-counts NodeStatus by 9 projects and HIDControls by 2. The ini files also spell dependencies three ways — bare names resolved via lib_extra_dirs, file:// paths, and registry packages — and put lib_deps on a shared [env_base] that each env extends.

Approach

Generate the map instead. .github/scripts/gen-platformio-impact-map.py walks both sources, takes the transitive closure, and either rewrites the map (--write) or fails with a diff when the committed copy disagrees (--check).

The generated file stays committed, so the graph is reviewable in a diff. The comparison is exact, not a superset check, so over-listing can't survive either — it drops CANProtocol -> Templates/SimGateway, which nothing supports.

The check runs in the existing node-ids job: pure Python, already on PRs, so drift is caught on every PR at no added cost. full_build_paths gains the generator, because a change to the derivation invalidates every selection.

Verification

Run with the selector (--mode affected --changed-file …), not by reading the JSON:

changed	before	after
Libraries/A4EC/…	2 (silently wrong)	9
Libraries/NodeStatus/…	full 33 (unknown library)	29
Libraries/CANProtocol/…	full 33 (missing project)	28
Libraries/HIDControls/…	full 33	31
Libraries/STM32Board/…	full 33	29
Libraries/PanelGroup/…	full 33	23
Libraries/DrumDisplay/…	full 33 (no key)	2
Libraries/PanelBridge/…	2	3
Libraries/SimGateway/…	1	3
Panels/Center_Armament/src/main.cpp	1	1
Libraries/PanelGroup/README.md	0	0
the map or either script	full 33	full 33

Drift detection tested both ways — a hand-edited entry and a newly added project each make --check exit 1 with a diff; the generator is idempotent on re-run; --mode full still enumerates all 33 projects.

Interaction with #339 — tested, not assumed

#339 hand-edits this same file (SoakRig + DrumDisplay entries), so the two conflict. I merged it into this branch in a throwaway worktree to check:

• the conflict is real but confined to .github/firmware-platformio-impact.json — nothing else;
• resolving it is git checkout --ours on that file, then --write. The generator now prints exactly that when it meets an unresolved conflict, instead of a json.loads traceback;
• both SoakRig boards are then picked up automatically from their own inis — no hand entries. Post-merge counts over 35 projects: A4EC 11, CANProtocol 30, DrumDisplay 3, HIDControls 33, NodeStatus 31, PanelBridge 3, PanelGroup 25, STM32Board 31, SimGateway 3.

That exercise also turned up a gap worth noting: #339 adds SoakRig under A4EC, CANProtocol, HIDControls, PanelGroup and STM32Board, but not NodeStatus, which both boards also pull in via CANProtocol. Nothing breaks if #339 merges alone — it just keeps under-listing until this regenerates on top. Merge #339 first.

Notes for review

• Two entries that looked wrong are correct and are kept: HIDControls -> Tests/CANProtocol and Tests/STM32Board. Both declare CANProtocol, which depends on HIDControls. Reading only their platformio.ini makes them look spurious.
• The firmware build job still does not run on PRs (if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'), and --mode affected is invoked nowhere in the repo. So this PR makes the map correct and drift-proof but does not by itself change what CI builds. Re-enabling affected-mode PR builds is a separate call and deliberately not made here.
• The node-ids job keeps its id and display name so any required-status-check config stays valid, even though it now does two checks.