Separate Edge-Case Semantics From Reused Happy-Path Logic

Summary#

Logic that is valid for the forward path, default tensor type, or common object shape often becomes wrong when reused in adjacent code paths. Handle derivatives, dynamic dimensions, low-precision dtypes, and alternate mask/object representations as first-class cases with their own semantics and regression tests.

Problem#

Implementations reused assumptions like "zero input implies zero gradient", "all masks expose ndim", or "QR works for every dtype", causing silent math errors or crashes in less common but supported modes.

Solution#

Define behavior per semantic path instead of reusing the nearest existing guard, branch on capabilities rather than nominal type shape, and add targeted tests for dynamic shapes, alternate object types, and precision-specific behavior.

Failure Modes#

Patching only the reported case can leave other adjacent edge cases inconsistent.
Branching on concrete classes instead of capabilities creates another round of brittle special cases.
Missing regression tests lets future refactors reintroduce the same assumption leak.

Sources#

https://github.com/tensorflow/tensorflow/pull/109358
https://github.com/tensorflow/tensorflow/pull/123015
https://github.com/tensorflow/tensorflow/pull/119869
https://github.com/huggingface/transformers/pull/46842
https://github.com/huggingface/transformers/pull/47381
https://github.com/huggingface/transformers/pull/47382
https://github.com/huggingface/transformers/pull/46854
https://github.com/huggingface/transformers/pull/47252
https://github.com/huggingface/transformers/pull/47369
https://github.com/huggingface/transformers/pull/47218
https://github.com/huggingface/transformers/pull/47032
https://github.com/huggingface/transformers/pull/47150
https://github.com/huggingface/transformers/pull/46603
https://github.com/huggingface/transformers/pull/47141
https://github.com/microsoft/ML-For-Beginners/pull/1002
https://github.com/microsoft/ML-For-Beginners/pull/1001
https://github.com/microsoft/ML-For-Beginners/pull/1000
https://github.com/microsoft/ML-For-Beginners/pull/994
https://github.com/microsoft/ML-For-Beginners/pull/991
https://github.com/microsoft/ML-For-Beginners/pull/990
https://github.com/microsoft/ML-For-Beginners/pull/989
https://github.com/ClickHouse/ClickHouse/pull/110779
https://github.com/ClickHouse/ClickHouse/pull/110397
https://github.com/ClickHouse/ClickHouse/pull/109529
https://github.com/ClickHouse/ClickHouse/pull/110849
https://github.com/ClickHouse/ClickHouse/pull/110519
https://github.com/ClickHouse/ClickHouse/pull/110688
https://github.com/ClickHouse/ClickHouse/pull/110864
https://github.com/ClickHouse/ClickHouse/pull/109901
https://github.com/ClickHouse/ClickHouse/pull/110491
https://github.com/ClickHouse/ClickHouse/pull/106724
https://github.com/ClickHouse/ClickHouse/pull/110598
https://github.com/ClickHouse/ClickHouse/pull/110528
https://github.com/vercel/turborepo/pull/13392
https://github.com/vercel/turborepo/pull/13391
https://github.com/vercel/turborepo/pull/13358
https://github.com/vercel/turborepo/pull/13382
https://github.com/vercel/turborepo/pull/13383
https://github.com/vercel/turborepo/pull/13380
https://github.com/vercel/turborepo/pull/13381
https://github.com/vercel/turborepo/pull/13378
https://github.com/vercel/turborepo/pull/13373
https://github.com/vercel/turborepo/pull/13374
https://github.com/vercel/turborepo/pull/13370
https://github.com/vercel/turborepo/pull/13369
https://github.com/vercel/turborepo/pull/13366
mined_at: 2026-07-17T23:22:26Z

Sagwan Revalidation 2026-07-18T00:00:30Z#

verdict: ok
note: 엣지 케이스별 의미와 회귀 테스트 원칙은 현재도 유효하다.

Sagwan Revalidation 2026-07-20T00:50:09Z#

verdict: ok
note: 일반 원칙 중심이며 최근 검증 후 바뀔 수치·권장안이 거의 없음