ci: emit per-board Kconfig dep graph as a release sidecar

[widgetii]

Summary

Adds `general/scripts/kconfig_graph.py` — a kconfiglib-based emitter that
walks the merged Buildroot tree, filters to currently-set + user-prompted
`BR2_PACKAGE_*` symbols, and writes two schema-versioned JSONs into
`IMAGES_DIR` alongside `sizes.json`:

• `kconfig-graph.-.json` — per-symbol type / prompt /
  depends / selects / selected_by (direct only)
• `kconfig-help.-.json` — per-symbol help text, split out

This feeds firmware-explorer v0.3's interactive configurator (separate PR)
which lets users toggle packages off, computes the cascade closure with
real Kconfig semantics, and downloads a defconfig fragment.

Why filter currently-set + user-prompted

• Buildroot's full `Config.in` defines ~7100 `BR2_PACKAGE_*` symbols across
  all upstream packages. Shipping all of them per board produces ~3 MB per
  platform with no UX benefit — the configurator can't usefully help you
  enable a package whose dependencies aren't already on this board.
• Non-prompted symbols (`_ARCH_SUPPORTS`, `SUPPORTS`, `HAS*`) are
  feature-detection Buildroot derives automatically; they aren't
  user-toggleable.
• The currently-set ∧ prompted intersection is closed under Kconfig's own
  dep tracking: every symbol referenced from an emitted symbol's
  `depends_on` / `select` chain is also set, so cascade resolution in the
  explorer doesn't lose references.

Empirical on a fully-configured cv200lite tree:

Artefact	Symbols	Raw	Gzipped
`kconfig-graph.hi3516cv200-lite.json`	48 (12 cascade-required)	14 KB	1.8 KB
`kconfig-help.hi3516cv200-lite.json`	47 entries	9 KB	3.8 KB

`selected_by` precision (subtle)

Kconfiglib's `rev_dep` expression includes symbols from both the select
source AND its condition (`select X if Y` puts Y in the rev_dep). For the
cascade UX we want direct selectors only — otherwise depended-on symbols
look like spurious "blockers". The emitter pre-builds a
`target_symbol → [direct selectors]` index by iterating `sym.selects`
across the whole symbol set, then filters to the currently-set ones.

Kconfig env contract

Mirrors `size_report.py`: `TARGET_DIR` / `BR2_OUTPUT_DIR` / `IMAGES_DIR` /
`OPENIPC_SOC_MODEL` / `OPENIPC_VARIANT`. Sets defaults for the env vars
Buildroot's Kconfig top-level expects (`BR2_BASE_DIR`, `BR2_EXTERNAL`,
`BR2_EXTERNAL_GENERAL_PATH`, `srctree`). Two non-obvious ones:

• `CONFIG_=""` — Kconfiglib defaults to stripping a `CONFIG_` prefix
  from `.config` lines; Buildroot's symbols carry the `BR2_` prefix
  included in the symbol table, so any non-empty prefix here makes
  `load_config` silently skip every line as "malformed".
• `BR2_SKIP_LEGACY=YES` — skip `Config.in.legacy` parsing. It only
  generates renamed/removed warnings that we don't surface.

Makefile + CI

• New `kconfig-graph` target mirrors `size-report:122-132`.
• `make deps` appends `pip install --break-system-packages kconfiglib`
  (kconfiglib is a single-file pure-Python module, ~10k LoC, MIT).
• `.github/workflows/build.yml` gets a "Build kconfig graph" step
  after the existing "Build size report". Best-effort: kconfiglib
  install or graph emit failure emits a `::warning::` and continues, so
  an upstream kconfiglib regression cannot sink a working build.
• `${{env.KCONFIG}}` + `${{env.KHELP}}` added to all three upload
  steps with the same pattern that already promotes `SIZES`.

Tests

`general/scripts/tests/test_kconfig_graph.py` drives `build_report()`
against a synthetic Config.in fixture, asserting: type / prompt /
currently_set filtering; depends_on / selects / selected_by extraction;
direct-selector precision; help-text split. Run via
`python3 general/scripts/tests/test_kconfig_graph.py` — 2 unittest
cases, both green locally.

Cross-repo dependency

Once this lands and one nightly cycle attaches the JSON to release tags,
`OpenIPC/firmware-explorer` PR (separate) lights up the configurator tab.
Until then, that PR's prebuild logs "no kconfig assets found in any
retained tag" and the tab stays disabled with a tooltip — by design.

Test plan

• Synthetic fixture tests: `python3 general/scripts/tests/test_kconfig_graph.py` — 2 passes
• Real-tree smoke: `make kconfig-graph` against `output-cv200lite` — emits 48 symbols / 14 KB raw / 1.8 KB gzipped
• Cascade correctness: BUSYBOX shows `selected_by: [BR2_INIT_BUSYBOX]`; ZLIB shows `selected_by: [BR2_PACKAGE_DROPBEAR_OPENIPC]` (one direct selector each, not the false-positive 40-element list from including condition expressions)
• CI green on this branch
• After merge: nightly attaches `kconfig-graph..json` + `kconfig-help..json` to release tags alongside `sizes.*.json`

🤖 Generated with Claude Code