[ashhart]

What this changes

A saved native capabilities --json file currently crashes the evidence index because native schema 1 maps families to format arrays. The index now reads those advertised facts at unknown measurement grade and rejects invalid supplied input with a bounded error instead of an empty success or traceback.

• Issue: Fixes evidence_index: --capabilities fails on native schema-1 output #2702.
• Stacked on: main 79d3dc661fa8e9e01b5674c83996f0db9d3f06cd; two commits 53b8f0a8d3429849e982eaa168474e93a5266963 and a1af2836fb5df8b422430d8ca64761635221286a.
• atc overlap --json: 9 Nov 2026, base 79d3dc6. Current working-tree overlaps with ci: collect native Linux host checks and guard exclusions #2095 and qwen35: dense BF16 loading and native prompt logprobs #2020 merge cleanly. The ordinary account portfolio remains incomplete and reports unrelated existing PR conflicts. The ticket publication guard separately checks the actual worktree against all peers and returns complete/ready under ATC 0.9.3. No overlap was waived.

The captured native fixture produces 15 family-format facts, one backend and one chip. Root backends do not create inferred family/backend pairs. All advertised facts retain unknown grade and no measured commit or result. Legacy family lists remain supported. Supplied input is bounded by file, collection, fact and name limits; errors do not echo payloads or machine paths. An absent source and an available empty source remain distinct. Existing Linux/Windows host CI now explicitly runs the evidence-index tests.

Checks on this head

• Head commit: a1af2836fb5df8b422430d8ca64761635221286a, tree a1c058bf31e22ef2d077974cf5a3e0238ab21801.
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: 386 inherited problems in 6144 files on both trees, zero new findings. Both full checker exits are 1 because the inherited findings remain.
• Tests run, with the command and pass, fail and skip counts: Python 3.12/Linux x86_64, python3 -m unittest discover -s tools/tests -p 'test_evidence*.py' -v: 32 passed on this head, zero failed/skipped. python3 -m unittest discover -s tools/tests -p 'test_bench*.py' -v: 148 passed on the rebased current source, zero failed/skipped. python3 tools/pr_check.py --base origin/main: zero errors/warnings. git diff --check passed. After the queue advanced main, the PR was rebased cleanly; all five feature files and the native producer contract remain identical to the reviewed source. The composition peer found no new asks. The public shallow test snapshot's base commit and entire tree match the selected source exactly; the five candidate file hashes match this head.
• CI: all 8 applicable No GPU checks pass on this head, run 42710856706. The two scheduled-only nightly jobs are skipped by their workflow conditions. Linux/Windows native host tests, root pytest, macOS dev-tools, Linux/Windows tool tests, fmt and release-path checks are green.
• Not run locally: native builds, GPU/model checks and application suites. This read-only Python tool cannot reach inference or an engine build API. The configured hosted native/host checks pass as recorded above; no real device or checkpoint qualification is inferred.

The final 12-case regression suite was actually run against unchanged main before the fix. It fails on the captured native object, invalid-input/CLI behavior and input limits; legacy compatibility is a passing control. The candidate also had two JSON-number edge failures before bounded error handling. Fresh parallel source review then found valid native unknown-chip and XPU display-chip values rejected by the token validator. The final 14-case suite actually fails on 76fc751 with three reader errors and three CLI failures. The revised reader omits null chips and accepts bounded chip display labels while keeping other names strict. All 32 evidence tests pass on this head. Both source peers rechecked the exact revision and report no remaining findings; these reports do not represent independent GitHub-account approval. No tolerances or engine settings changed.

Receipt

n/a. This read-only evidence tool cannot reach inference.

• Main against branch: the captured native file and malformed-input regressions fail on main and pass on this head; no inference comparison applies.
• Boxes it reaches: host Python clients; CI covers Linux and Windows. No device dispatch changes.
• New switch? Default, accepted values, profile row: n/a; the existing optional --capabilities argument is unchanged.
• Environment, with the TensorFold commit, Zig and backend toolchain, machine and GPU, checkpoint and revision: source head/base above, Python 3.12/Linux x86_64; Zig/backend/GPU/checkpoint n/a for this host tool.
• Paths and platforms the change reaches, which I ran and which I could not: Linux host suite passed; Windows tool and all configured hosted host results now pass. No local device run.
• Exactness, from tools/bench_concurrent.py --alone --serial or token_sha against "draft": false: n/a; inference is unreachable.
• Main against this branch, same machine and checkpoint, posted before asking for a run on our machines: n/a; no model run requested.
• Decode speed before and after, from tools/bench_openai.py: n/a; no decode changes.
• Prompt speed before and after: n/a; no prompt processing changes.
• Licence of any new weights or drafter: n/a; none added.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• Regression tests fail before the change and pass after it; legacy controls pass. Standard-library tests need no optional dependency.
• Added comments and docstrings are one line. No measurements or history added to source comments.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines. Entire source tree, hidden/ignored files, reachable history and outgoing identity inspected.

[ashhart]

/atc exchange a86b77f5f3dbf11ee5e6052a6e73bc0d89b4011d707a9bd1b2964e6e711e4e68