[inureyes]

Summary

• Disable LTO in [profile.test-fast] so macOS workspace integration-test binaries no longer enter the ThinLTO link path that produced missing internal Rust symbols.
• Preserve the optimized test behavior that matters for MLX numerics: opt-level = 3, codegen-units = 16, incremental compilation, unstripped local test binaries, and inherited unwind behavior remain unchanged.
• Leave [profile.release] unchanged with fat LTO and one codegen unit, and update Cargo, Makefile, contributor, installation, and nightly guidance to document the test/release trade-off accurately.
• Add bilingual technical reports covering the reproduction, decision, validation evidence, and residual release-codegen gap.

Root cause

The full workspace no-run graph repeatedly failed while linking large integration targets such as molmo_parity and qwen3_omni_moe_parity, with unresolved LLVM-suffixed Rust drop-glue symbols. The failure reproduced with CARGO_BUILD_JOBS=1, while a targeted integration-test build could pass under a smaller unit graph, isolating the failure to the explicit ThinLTO workspace link shape rather than the test source or concurrent Cargo link jobs.

Validation

• TOML profile assertions: [profile.test-fast] has lto = false, opt-level = 3, and codegen-units = 16; [profile.release] remains lto = true, codegen-units = 1, strip = true, opt-level = 3, and panic = "unwind".
• cargo test --profile test-fast --features metal,accelerate --test molmo_parity --no-run
• cargo test --profile test-fast --features metal,accelerate --test qwen3_omni_moe_parity --no-run
• cargo test --workspace --profile test-fast --features metal,accelerate --no-run — passed; all workspace test targets linked in 9m19s on the cold corrected profile.
• cargo test --workspace --profile test-fast --features metal,accelerate --no-fail-fast -- --test-threads=1 — passed with zero failures across all workspace test and doctest binaries.
• cargo clippy --workspace --all-targets --features metal,accelerate -- -D warnings
• cargo fmt --all -- --check
• git diff --check origin/main...HEAD
• GitHub CI: formatting, clippy, dependency policy, crate versions, kernel dtype keys, cross-repository references, OpenXLA feature compilation, and CLA checks passed before the report-only commit; the refreshed checks are expected to exercise the same implementation tree.

Review

• Independent correctness review: no findings.
• Security/performance review: no findings at any severity; the intentional test-only LTO trade-off and unchanged release workflow were verified.
• Finalization review: scope, closing link, profile isolation, documentation, and focused link checks are complete with no blockers.

Reports

• TECHNICAL_REPORTS/1407-disable-test-thinlto-20260825.en.md
• TECHNICAL_REPORTS/1407-disable-test-thinlto-20260825.ko.md

Closes #1406

[inureyes]

Security/performance review

• Findings: no CRITICAL, HIGH, MEDIUM, or LOW security/performance findings in the changed build-profile, Makefile, workflow, or documentation files.
• Build-profile audit: [profile.test-fast] now overrides inherited release LTO with lto = false while retaining inherits = "release", opt-level = 3, codegen-units = 16, incremental = true, and strip = false; [profile.release] remains fat-LTO with codegen-units = 1 and is unchanged by this PR.
• CI/performance trade-off: the change intentionally removes test-time cross-crate LTO from the workspace gate to avoid macOS ThinLTO linker failure and should reduce test-link pressure; the documented residual risk is limited to defects that reproduce only under release LTO or single-codegen-unit release codegen, with release.yml and manual cargo test --release --features metal,accelerate still covering that escape hatch.
• Validation: TOML profile assertions passed, git diff --check passed for changed files, .github/workflows/release.yml has no PR diff, and cargo metadata --no-deps --format-version 1 parsed successfully. Broad workspace Cargo gates were intentionally not rerun here because the orchestrator already owns them.