Fix board_capacity dropping known values on partial override

[divemba]

Summary

• board_capacity() in ebuild/build/footprint.py discarded the reference-table
  value for whichever of flash/ram a board YAML didn't explicitly override,
  instead of falling back to it. A board config that set only ram_size, for
  example, would report flash as unknown even though the board name matched
  an entry in _REFERENCE_CAPACITY.
• Missing fields are now backfilled from the reference table when at least one
  field is overridden, so partial overrides no longer lose known data.
• Removed a duplicate pair of marker declarations (cross_repo, qemu) in
  pytest.ini, left over from a prior merge.

Test plan

• Added test_a_partial_override_fills_the_rest_from_the_reference_part
  covering both partial-override directions (flash-only, ram-only)
• pytest -q — 562 passed
• Independent code-review pass — no findings

[srpatcha]

Review — ebuild#105 "Fix board_capacity dropping known values on partial override"

head: acb7d34 author: divemba ci: none reported (checks.txt is 0 bytes)

Verdict: The bug is real and the fix is right for the case it describes — I reproduced both. What it also does, unintentionally and untested, is make a malformed or explicitly-zero size override indistinguishable from an omitted one, so a typo in a board YAML now silently resolves to the reference part's figure. Separately, the function has been re-indented from four spaces to three, which turns a six-line change into a nineteen-line rewrite.

Findings

#	Severity	File:line	Finding	Recommended fix
1	Medium	ebuild/build/footprint.py:213-226	The backfill cannot tell "not specified" from "specified wrongly", so a typo now reads as a valid capacity. _as_bytes() (footprint.py:231-241) returns None for three different situations: the key is absent, the value does not parse (int(str(value), 0) raising ValueError), and the value parses to zero or less. Before this change all three produced (None, ram) — unknown, which is what the reporting path treats as "no percentage". After it, all three are backfilled from _REFERENCE_CAPACITY. Executed against both trees with board="stm32f4": flash_size: "1MB", ram_size: 327680 → master (None, 327680), this PR (1048576, 327680) flash_size: 0, ram_size: 327680 → master (None, 327680), this PR (1048576, 327680) ram_size: 327680 (the intended case) → master (None, 327680), this PR (1048576, 327680) Three semantically different board YAMLs, one answer. 1MB is not a hypothetical typo — the field accepts 0x40000 and plain integers, so a unit suffix is the obvious thing to get wrong, and the developer's own new test uses the string form. The consequence is not a crash: ebuild prints a flash percentage computed against 1 MiB that the developer never wrote and cannot see in their config, which is the opposite of §9.2's "actionable diagnostics with remediation guidance" and undermines the "per-target flash/RAM reports" the same section promises. flash_size: 0 meaning "this variant has no on-board flash" gets the same treatment.	Distinguish the cases at the source. _as_bytes() returning None for a malformed value is the conflation; have it raise, or return a sentinel, so board_capacity() can backfill only genuinely-absent keys. The smallest version that keeps the fix and closes the hole: check key presence rather than parsed truthiness — if "flash_size" not in memory and reference: flash = reference[0], same for ram_size — which backfills exactly the omitted field and leaves a bad value as unknown. Add both cases to the new test; they are two lines each.
2	Medium	ebuild/build/footprint.py:203-226, tests/unit/test_footprint.py:156-165	The function is re-indented from four spaces to three, and the new test method from eight to seven. Every line of board_capacity moves: the continuation of the signature drops from column 19 to 18, the docstring and the whole body go to three-space indent. Python accepts it — I confirmed py_compile is clean and the tests pass — but footprint.py has 175 lines at four-space indent and, after this patch, 9 at three, all of them from this diff. The cost is on review and on history: a six-line logic change is presented as 23+ 19-, so a reviewer has to diff the whitespace mentally to find the actual change, and the next black or ruff format run over this file will produce a spurious commit touching code nobody edited. The same slip is in the test — the method body sits at seven spaces inside a class whose other methods use eight.	Re-indent to four spaces throughout, both files. The logic change is then five added lines and reads as one. Nothing else needs to move.
3	Medium	— (checks.txt is empty; PR body "Test plan")	No CI check of any kind ran, and two of the three checklist items are claims with nothing behind them. checks.txt is 0 bytes; createdAt and updatedAt are both 2026-09-02T00:52:43Z, opened and never touched, while ebuild#103 and #104 in this same batch carry 24 and 30 checks each. So pytest -q — 562 passed is a local run with nothing independent behind it, and "Independent code-review pass — no findings" names no reviewer, no tool and no output — under the brief's rule that an unsupported "verified" is itself the finding, that line is the finding. The count is also off: 562 is the number of tests collected, not passed. On this host, pytest -q against the patched tree gives 1 failed, 558 passed, 3 skipped — and to be fair to the author, that one failure is tests/ebuild/test_build_dir_resolution.py::test_end_to_end_build_from_outside_produces_the_binary failing with No module named ninja, which is my environment and is pre-existing on master (1 failed, 557 passed, 3 skipped unpatched). With ninja installed the tree would give 559 passed and 3 skipped, so the honest figure is 559, not 562.	A maintainer approves the workflow runs. Replace "Independent code-review pass — no findings" with either the reviewer's name or nothing, and quote the real pytest tail rather than a collected count.
4	Low	ebuild/build/footprint.py:206-211	The docstring now describes the old contract. It still reads "A project's own board YAML wins over the reference table", which was accurate when any override suppressed the table wholesale and is not accurate now — the YAML wins per field, and omitted fields fall back. That sentence is the only specification this function has, and the whole point of the PR is that the previous all-or-nothing reading was wrong. A stray blank line was also added inside the docstring, and there are two blank lines in the middle of the new test method's body.	Rewrite the sentence: "Each of memory.flash_size and memory.ram_size overrides the reference table independently; a field the YAML omits falls back to the table entry for board." Drop the stray blank lines.
5	Low	pytest.ini:20-21	The duplicate cross_repo and qemu marker declarations are genuinely duplicated on master and removing them is correct. It is also unrelated to the title, which describes only the board_capacity fix — the body does mention it, so this is not a hidden change, just one the commit history will not explain.	Keep it; split the commit so git log --oneline shows both. No code change.

Verified clean, executed against origin/master extracted with git archive into /tmp/eb105 with this patch applied:

• The bug is real and the fix works for the case it targets. On master, board_capacity("stm32f4", {"memory": {"ram_size": 327680}}) returns (None, 327680) — the reference flash figure for a board whose name matches _REFERENCE_CAPACITY is discarded because the config mentioned RAM. With the patch it returns (1048576, 327680). That is exactly what the summary claims.
• tests/unit/test_footprint.py is 44/44. The two new assertions check out arithmetically against _REFERENCE_CAPACITY at footprint.py:49-58: "stm32f4": (1024 * 1024, 192 * 1024), so flash-only 0x40000 → (262144, 196608) and ram-only 320 * 1024 → (1048576, 327680). Both directions are covered, which is more than the summary needed to do.
• The restructuring is behaviour-preserving where it should be. Hoisting the _REFERENCE_CAPACITY.get() lookup above the board_config branch does not change the no-config path — test_a_board_config_without_memory_falls_back and the surrounding cases still pass — and the board is None guard is preserved by the conditional expression.
• It compiles. py_compile clean on both changed Python files despite the indentation (finding 2 is about style and review cost, not correctness).
• The pytest.ini lines really were duplicated on master, at the two lines this patch removes.

Architecture conformance

Conforms. §21 Tier 1 — Foundation (ebuild). ebuild/build/footprint.py is host-side tooling: no import, link line or manifest entry crosses a tier, and §5.1's "eBuild understands the complete graph but is not a runtime dependency" is unaffected — nothing here is compiled into an image. tests/unit/ is the right home per .ai/architect.md.

The design section this actually serves is §9.2's "per-target flash/RAM reports" and §25.2's "debugger, trace, memory/flash reporting" as an MLP requirement. That is also why finding 1 matters more than its size suggests: a capacity figure that silently substitutes a reference part's number for a value the developer mistyped makes the report confidently wrong rather than usefully blank, and §28's evidence policy is explicit that a memory-footprint claim requires "exact feature configuration and linker output". A backfilled capacity is neither.

No proposal appended. The master design has nothing wrong or missing here — §9.2 and §28.1 already say what a footprint report owes the developer, and this is a repository-level defect against that, not a gap in the specification.

Proposed changes

1. Distinguish absent from malformed before backfilling, and cover both in the test (finding 1). This is the one I would ask for before merge; it is a few lines and it is the difference between the fix being right and being right-shaped.
2. Restore four-space indentation in both files (finding 2). Independent of everything else and makes the rest of the review cheap.
3. Get the workflow runs approved, and correct the two claims in the Test plan (finding 3).
4. Update the docstring to the per-field contract (finding 4).

Items 2–4 are independent of each other and of item 1.

Not checked

• Nothing ran in CI, so nothing here is independently corroborated. My runs are on this host with a uv-managed pytest that needed PYTHONPATH=/usr/lib/python3/dist-packages to import yaml, and without ninja — which is why one pre-existing end-to-end test fails for me and presumably not for the author. That is a weaker position than for #103 and #104 in this batch, both of which had full check runs.
• I did not check callers of board_capacity(). Finding 1 describes the function's contract changing; what a wrong capacity actually does downstream — whether Report merely omits a percentage, prints a misleading one, or gates anything — I inferred from test_percentages_appear_only_with_a_known_capacity existing, not from reading the reporting path.
• No board YAML in hardware/board/ was examined. Whether any real config in this repo or in eos currently sets one of the two fields, sets a malformed one, or sets flash_size: 0, I do not know — so finding 1's impact is demonstrated on synthetic inputs and its prevalence is unmeasured.
• The _REFERENCE_CAPACITY figures themselves were not checked against the parts. stm32f4 → (1 MiB, 192 KiB) is annotated STM32F407 and is plausible, but the accuracy of the table is the premise of both the old and new behaviour and I took it as given.
• No linter or formatter was run. Finding 2 rests on counting indentation in the file (175 lines at four spaces, 9 at three) and on PEP 8, not on a ruff or black run — I did not check whether this repository configures either, or whether CI would have flagged it.
• mergeStateStatus: BLOCKED, mergeable: MERGEABLE, reviewDecision: REVIEW_REQUIRED. No merge attempted. tests/unit/test_footprint.py also appears in #102's merged history and in #103's and #104's stacked diffs; those are already on master, so a conflict here is unlikely but the file is busy.
• The local ebuild checkout is dirty and was skipped by the sync step, and sits on branch v90. I read origin/master through git show and git archive; nothing in the working tree was touched, and all patched trees are under /tmp.

---

Automated architecture review of acb7d342badb — scheduled, model claude-opus-5, checked against the EmbeddedOS Master Design v2.0. Advisory only: this reviewer never approves, requests changes, or merges. Reply here to discuss or push back — a wrong finding is a bug worth reporting.