alint — tensorflow/tensorflow

1,185 textproto API goldens. Two discipline layers stacked. Zero existing tools that see the cross-language conventions. tensorflow demand-validates the v0.11+ `cross_language_implementation_complete` flagship at TWO topologies in one repo.

Why this matters

tensorflow/tensorflow is the canonical multi-language ML monorepo on GitHub (~190k stars). Every modern deep-learning library — TFX, TensorFlow.js, TensorFlow Lite, TF Hub, JAX-via-XLA — is substrated on this tree. It is the canonical “core + per-language bindings” topology: one Python frontend over a C++/CUDA runtime plus 5 per-language bindings (C, C++ via cc/, Java legacy, Go, JS codegen scaffolding) that mirror parts of the core API, PLUS a separate TFLite per-device frontend that itself has 4 per-language bindings (Python, Java, Swift, ObjC) sharing the same core C runtime.

No per-language linter sees the cross-language API parity surface. pylint only sees Python. buildifier only sees Bazel. clang-format only sees C++. The invariants this case study enforces — TFLite Swift Sources/<X>.swift ↔ Tests/<X>Tests.swift parity, Python TFLite <x>.py ↔ <x>_test.py parity, ObjC API headers ↔ tests parity, the 1,185 textproto goldens existing AND having the canonical marker, the 5 per-Python-version requirements locks all present and non-empty — are exactly the layer alint owns and nothing else does.

Headline catch

tensorflow has TWO discipline layers stacked: file-shape parity (every source has a paired test, by language convention) AND API-shape parity (every public Python symbol has v1+v2 textproto goldens; every binding mirrors the core API). alint covers Layer 1 cleanly today via pair/for_each_* primitives. Layer 2 — the cross-language API mirroring — is exactly the v0.11+ cross_language_implementation_complete rule kind, and tensorflow is the cleanest single-repo demonstration of the pattern.

The pitch lands harder when paired with the TFLite Swift finding. A pure file-shape parity pass, expressible today, surfaces:

**5 of 11 TFLite Swift Sources/.swift files have no matching Tests/Tests.swift (CoreMLDelegate, Delegate, InterpreterError, SignatureRunnerError, SignatureRunner)
4 ObjC API headers have no matching tests/ partners
The deeper Python ↔ Java ↔ Swift ↔ ObjC ↔ Go ↔ JS API symbol parity is enforced today by tribal knowledge + code review, not static check

No Swift tool catches the parity drift because no Swift tool sees the parity convention from above.

Where alint earns its keep here

The TWO discipline layers stacked are the v0.11+ flagship demonstration. apache/arrow gives the all-peers topology; TF gives the core+bindings topology. The v0.11+ cross_language_implementation_complete primitive needs to handle BOTH topologies, and TF surfaces both in one repo (per-source ↔ per-test within one language; core ↔ N bindings across languages). Saturated demand: arrow + tensorflow + protobuf + angular + flutter — 5 confirmed sources, the v0.11+ flagship ship-target, with TF’s 1,185 textproto API goldens as the canonical Python-side example.
TFLite per-language file-shape parity is expressible TODAY. 3 declarative rules (tensorflow-lite-swift-source-has-test, tensorflow-lite-objc-api-has-test, tensorflow-lite-python-source-has-test) catch what no Swift / ObjC / Python tool sees alone — the cross-language convention that every binding follows the same discipline.
requirements_lock_3_*.txt matrix is the 8th confirmation for cross_file_value_equals. Past saturation; v0.10 must-ship.
opensource_only.files is the 6th confirmation for registry_paths_resolve. Past saturation; v0.10 ship-target.
API goldens regen-and-diff is the 5th confirmation for generated_file_fresh. Past saturation; v0.10 ship-target.
Apache 2.0 + GHA hardening + OSS-baseline bundles apply cleanly. Only one custom rule extends the Apache header check to BUILD/.bzl files, which the bundled rule’s default file- extension list doesn’t cover.

Future story angles

Second polyglot tile on alint.org/examples (after apache/arrow). Same framing as arrow — for cross-language monorepos, no per-language linter sees the cross-language conventions; alint is the layer that does — but at the core+bindings topology, generalising to every multi-binding spec (TFLite, TFX, TF Hub, JAX/XLA bindings, gRPC, envoy, protobuf-codegen, angular goldens, flutter embedders).
v0.11+ flagship launch narrative — pair TF’s 1,185 textproto goldens with arrow’s per-language schema parity, protobuf’s 10 in-tree language bindings, angular’s 50 API goldens, and flutter’s 6 native-OS embedders for a “5 saturated sources” launch-post on cross_language_implementation_complete. TF is the cleanest demonstration because it shows BOTH topologies in one tree.
bazel-monorepo@v1 bundled ruleset draft — TF’s BUILD-file presence + *.bzl Apache-header + MODULE.bazel/WORKSPACE/ BUILD triad checks recur in any Bazel monorepo (TF + grpc + envoy + many internal Google + Pinterest + Lyft repos). Packaging these as a bundled ruleset would let alint claim the “Bazel-tier polyglot monorepo” niche; TF is the canonical example to author it against.
alint suggest detector breadth narrative — current suggester surfaced only oss-baseline + agent-hygiene (medium) against the live TF tree; the bundled-ruleset detector doesn’t yet recognise the Bazel monorepo shape. v0.10+ suggester-improvement candidate alongside the bundled ruleset.