fix(deps): tract 0.23 to clear RUSTSEC-2026-0217 (fixes #657)

[1jehuang]

Clears RUSTSEC-2026-0217, the last remaining CI failure once #663 lands.

Stacked on #663. Merge that first.

Why 0.23 rather than a patch bump or an ignore

The prior investigation on #657 established that the in-line fix is unreachable, and I confirmed it: tract-data 0.21.16 and 0.21.17 pin half =2.4.1 exactly, while naga (via vello in jcode-desktop2) requires half ^2.5. Neither direction resolves without dropping the desktop2 GPU stack. The 0.23 line drops that pin.

I went with the upgrade rather than a scoped ignore because the exposure is not the comfortable "we only parse bytes we shipped" case. load_from_dir downloads model.onnx from huggingface.co when it is absent and then parses the fetched artifact, so the vulnerable NNEF parser does run over data retrieved at runtime.

The bump is mechanical

jcode-embedding is the only consumer. Three renames, all compiler-directed:

• SimplePlan<TypedFact, Box<dyn TypedOp>, Graph<..>> becomes Arc<RunnableModel<TypedFact, Box<dyn TypedOp>>>. 0.23 dropped SimplePlan's third (model) type parameter, and into_runnable now returns the Arc<RunnableModel<..>> alias directly.
• TValue::to_array_view becomes to_plain_array_view, 2 call sites.

Diff is 3 files: the version bump, those renames, and Cargo.lock.

Verification

• scripts/security_preflight.sh --strict passes. cargo tree -i tract-nnef reports 0.23.4 and the advisory is gone. Only the 11 pre-existing allowed warnings remain.
• Inference is correct, not merely compiling. This matters more than the compile: a silently-wrong embedding would pass a build. The local MiniLM model happens to be present on this host, so minilm_related_beats_unrelated_if_present actually executed rather than taking its skip path, and its cosine-similarity assertion against the real model still holds. cargo test -p jcode-embedding gives 5 passed, and the --ignored latency probe runs real inference end to end (model load 164ms, 380ms per embed).
• Full scripts/check_guardrails.sh: all gates pass, no baseline changes.
• cargo metadata --locked is clean.

docs/SECURITY_DEPENDENCIES.md records the resolution and the half conflict, so nobody repeats the 0.21.16 attempt.

Worth your judgement

This is a major version bump of an inference dependency, which is why #657 was labeled autonomous: no. I am putting it up because it turned out to be three renames rather than a real migration, and because the alternative was an ignore over a runtime-downloaded artifact. But the semantic check that gives me confidence depends on a locally-present model, and that test skips silently on a machine without it, including CI. So the quality evidence is from my machine, not from the pipeline. If you would rather ship the ignore and schedule the upgrade deliberately, that is entirely reasonable and this branch can just be closed.

Separately: that test skipping rather than failing when the model is absent is the same "gate that looks green because it never ran" pattern as the warning budget in #663. Probably worth making the model a CI fixture, but that is out of scope here.

--- — Jcode agent (automated triage), on behalf of @1jehuang

[greptile-apps]

Greptile Summary

This PR upgrades local embedding inference to tract 0.23.4 and adds guardrails around the migration. The main changes are:

• jcode-embedding now depends on tract-hir and tract-onnx 0.23.
• Embedding and reranker output handling now uses the tract 0.23 array-view API.
• CI fetches the MiniLM model and fails if the numeric-stability test skips.
• Secret-input prompts now echo masking characters and have PTY regression tests.
• Additional guardrails cover missing Rust module files, TUI copy tests, and login helper extraction.

Confidence Score: 5/5

Safe to merge with low risk after the stacked base lands.

No blocking issues were found. The tract bump is localized to jcode-embedding, and CI now forces the MiniLM numeric-stability test to run with fetched fixtures instead of silently skipping.

Files Needing Attention: No files require special attention.

T-Rex Logs

What T-Rex did

• Validated the dependency resolution for jcode-embedding and confirmed tract-nnef v0.23.4 is the locked reverse dependency; the command exited with code 0.
• Executed the embedding crate tests with nocapture; six unit tests passed, and MiniLM tests were skipped due to missing local model.
• Ran the latency probe for embedding; the probe reported a missing model directory and that the model was not installed, then skipped, exiting with code 0.
• Reached the strict preflight step and attempted cargo-audit, but cargo-audit is not installed, resulting in exit code 1.

_{Ran code and verified through T-Rex}

Important Files Changed

Filename	Overview
.github/workflows/ci.yml	Adds CI coverage for missing module files, PTY secret-input tests, embedding numeric-stability with fetched MiniLM fixtures, and copy-badge TUI tests.
Cargo.lock	Updates the locked dependency graph for the tract 0.23.4 upgrade and associated transitive changes.
crates/jcode-embedding/Cargo.toml	Bumps tract-hir and tract-onnx from 0.21 to 0.23.
crates/jcode-embedding/src/lib.rs	Adapts the embedding and reranker code to tract 0.23 APIs and adds a numeric-stability regression test.
crates/jcode-base/src/secret_input.rs	Adds masked keystroke feedback for raw-mode secret entry while preserving non-TTY fallback behavior.
crates/jcode-base/src/secret_input_pty_tests.rs	Adds Unix PTY tests proving secret input echoes masks, handles backspace, and does not reveal the secret.
scripts/check_module_files.py	Adds a tracked Rust source scanner that fails when mod x; lacks a resolvable module file.
src/cli/login.rs	Splits login helper text into focused modules and wires existing-key notices and local endpoint next-step hints into the OpenAI-compatible flow.
src/cli/login/existing_key_notice.rs	Adds testable notice generation for already-configured OpenAI-compatible API keys.
src/cli/login/next_step.rs	Extracts local endpoint next-step guidance and tests provider-specific hints.

Sequence Diagram

sequenceDiagram
participant CI as GitHub Actions CI
participant HF as Hugging Face
participant Test as jcode-embedding tests
participant Tract as tract 0.23 runtime

CI->>HF: Download model.onnx and tokenizer.json
Test->>Tract: Load ONNX with tract_onnx 0.23
Test->>Tract: Run MiniLM embedding inference
Tract-->>Test: Return f32 tensor via to_plain_array_view
Test->>Test: Assert dimensions, norm, and reference components
Test-->>CI: Fail if skipped or numerically drifted

Loading

_{Reviews (6): Last reviewed commit: "fix(deps): move jcode-embedding to tract..." | Re-trigger Greptile}

[1jehuang]

Ran the check that actually mattered for this bump and had not been done: do tract 0.21 and 0.23 produce the same embedding numbers?

This is not covered by "it compiles" or by minilm_related_beats_unrelated_if_present. Persisted memory embeddings are keyed by model_id, which stays all-MiniLM-L6-v2 across a tract bump, so the model-mismatch filter in memory.rs:854 cannot notice a numeric change. If 0.23 embedded text differently, every stored vector would silently become incomparable with freshly computed ones, and the related-vs-unrelated assertion would still pass because both vectors move together.

Measured, on both versions, same model file

Ran an identical probe on fix/issue-662-ci-red (tract 0.21.10) and this branch (0.23.4):

text="how do I set the cargo build profile"
  dim=384 norm=1.000000
  head=[0.040730, 0.103227, 0.065710, 0.032524, -0.012075, 0.005709, 0.029490, -0.038528]

Across all three probe texts and 8 reported dimensions, exactly one value differs:

0.21.10:  ... 0.028180 ...
0.23.4:   ... 0.028181 ...

A single 1e-6 delta, i.e. float32 rounding. Worst case if every one of the 384 dimensions drifted that much: |error| ≈ 2.0e-5, cosine deviation from 1.0 of about 1.9e-10. That is roughly 10^8 times smaller than any retrieval threshold in the codebase. Existing persisted embeddings stay compatible; no re-embedding is needed.

Landed as a test, calibrated by mutation

minilm_embedding_is_numerically_stable_across_inference_engines pins dimensionality, L2 norm, and the first 8 components against the 0.21.10 reference at 1e-4 tolerance. Verified it is neither vacuous nor over-tight:

Condition	Result
tract 0.21.10	passes
tract 0.23.4	passes
non-uniform perturbation injected into the pooled output	fails: dim 0 drifted beyond float32 rounding: diff 8.3e-4

So it tolerates the real rounding and catches a change three orders of magnitude smaller than one that would actually degrade retrieval. It sits on #663 rather than here, so the reference values are pinned on the pre-bump engine and this PR has to satisfy them.

One honest limitation, unchanged: like its neighbours, the test skips when the MiniLM model is absent, which includes CI. Making the model a CI fixture is the fix, and is worth doing separately.

--- — Jcode agent (automated triage), on behalf of @1jehuang