fix: uninitialized input matrix with thread <= 10

[bact]

• uniform() filled only thread/10 of the input matrix, plus no tail when thread=10; the rest stayed uninitialized -> NaN or non-deterministic training for thread<=10
• Fill all 10 blocks + tail for any thread count (sequentially for thread=1). Values for thread>=11 are unchanged; thread<=10 now gets a fully initialized matrix, so its output differs from before
• Applies to both random and pretrained-vector init

kUniformBlocks lets uniformThread and uniform share the block count.

The test runs _CHILD (a small script that imports fasttext) in 5 fresh processes and checks, for random and pretrained init across threads 1, 2, 10, 11, 12, that every value is initialized and the matrix is identical for every run.

[bact]

Before this PR (uses runs from PR 13 as examples):

• test (ubuntu-latest, 3.9) - 45 passed, 38 skipped
• test (windows-latest, 3.13) - 40 passed, 43 skipped
• test (macos-latest, 3.14) - 72 passed, 11 skipped

With this PR:

• test (ubuntu-latest, 3.9) - 80 passed, 4 skipped
• test (windows-latest, 3.13) - 80 passed, 4 skipped
• test (macos-latest, 3.14) - 80 passed, 4 skipped

[Copilot]

Copilot review overview

🟡 Changes recommended

A critical seed-overflow issue and several regression-test gaps remain unresolved.

Review effort: Lite
Findings: 1 · 1

Open (2)

• Signed seed addition can overflow in uniformThread · New
• Regression test omits threaded scheduling paths · New

What changed in this PR

Fixes DenseMatrix::uniform so the input matrix is fully initialized for any thread count, with regression coverage.

Changes:

• Distributes all matrix blocks across workers.
• Shares block configuration between initialization paths.
• Adds deterministic Python regression tests.

File	Summary	Review findings
src/​densematrix.cc	Initializes every matrix block across thread counts.	Critical: avoid signed overflow when deriving RNG seeds from INT32_MAX.
python/​fasttext_module/​fasttext/​tests/​test_matrix_init.py	Adds initialization and determinism coverage.	Moderate: allow valid zero values, use a small explicit bucket count, and cover threaded cases beyond thread=1.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[bact]

Thank you for the review. I will look at the comments and suggestions.

[bact]

Fixed issues as suggested.

• Ubuntu - 84 passed, 4 skipped
• macOS - 84 passed, 4 skipped
• Windows - 84 passed, 4 skipped

(the additional 4 passed are from the recent PRs merged)

[Copilot]

Copilot review overview

🟡 Changes recommended

The claimed pretrained-vector initialization path is not covered by the regression test.

Review effort: Lite
Findings: 1 · 1

Open (2)

• Regression test omits pretrained vector initialization coverage · New
• Child process determinism test runs only three repetitions · New

Resolved since last review (2)

• Signed seed addition can overflow in uniformThread
• Regression test omits threaded scheduling paths

[Copilot]

Copilot review overview

🔵 Needs a closer look

The test assertion does not fully validate the claimed determinism behavior.

Review effort: Lite
Findings: None

Resolved since last review (2)

• Regression test omits pretrained vector initialization coverage
• Child process determinism test runs only three repetitions

[Copilot]

Copilot review overview

🟢 Approval recommended

Only a minor, non-blocking test nit remains.

Review effort: Lite
Findings: None