fix(bazel): Cargo feature parity and Windows builds

[ScriptedAlchemy]

Bazel builds were resolving different Rust features and test runtime inputs from the existing workspace. That changed canonical JSON identity digests, omitted grammars and test-transport behavior, and left test subprocesses and fixtures inaccessible.

This change corrects target-side serde_json features, shares the workspace test feature resolution, and generates test binaries with the features their suites require. Test launchers stage runtime resources, provide absolute executable and toolchain paths, retain canonical test crate names, and isolate each test's home and profile. Installed host configuration prefers the stable PATH executable over a Bazel build output.

Windows fixes shorten external build-script paths, patch the aws-lc header include, and apply platform configuration after generic defaults so the Windows workspace-status command and UTF-8 compiler flags actually take effect. CI queues are scoped to each PR instead of blocking unrelated validation globally. The macOS filename probe proves ordinary writes succeed before recognizing an unsupported filename.

Validation through Bazel:

• Original domain suite: 169 tests passed, including strict JSON identity assertions and runtime fixture reads.
• Focused macOS churn tests: 2 passed under the sandbox.
• Runtime environment probe, private-fs, and framing targets passed.
• Generated BUILD validation, repository gates, dashboard, and feature gates passed on earlier shared-fix commits.

Full CI is running on the current head. Windows launcher and dependency-resolution failures still need confirmation from the diagnostic run; this draft is not yet green across all platforms. Temporary Windows diagnostic output will be removed after the failures are resolved.

[changeset-bot]

⚠️ No Changeset found

Latest commit: 58fa8e2

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[ScriptedAlchemy]

Windows lane status after 558b65f (run 37438541376):

• Fixed by the MAX_PATH change: tokensave-large-treesitters' build script now compiles.
• Still failing, addressed in f46beb2:
  • //:product_git_sha: the workspace status wrapper now reads git directly, because a bare bash can resolve to WSL.
  • //dashboard:bundle_sha256: Windows genrules get a stripped PATH with no Python. They now get the runner's PATH, and the genrule uses the first interpreter that starts.
• Still failing, cause not yet known:
  • aws-lc-sys build script: its 6 MB of stderr was cut off by --experimental_ui_max_stdouterr_bytes. CI now keeps up to 16 MB and passes --verbose_failures, so the next run will show the real error.
  • tracedecay-code-index and tracedecay-lsp: every variant fails with E0463 can't find crate for all of its dependencies, members and externals alike. Other crates compile. Both crates depend directly on tracedecay-code-extraction. I'm waiting on the verbose rustc command from the next run.
• Test failures still open:
  • serde_json preserve_order leaks from tree-sitter's build-dependency into the Windows target build. This changes JSON key order and identity digests.
  • Timing-sensitive tests under load.

A new workflow_dispatch run with run_os=true is queued on f46beb2.

---

Generated by Claude Code

[ScriptedAlchemy]

Test Windows failed on 14b8ffd4 (run 37511829834): 53 tests passed, 12 failed and 95 failed to build. I addressed it in db0abfdb.

• //:product_git_sha. Bazel runs the workspace status command through cmd.exe, which reads scripts/bazel/... as the command scripts followed by the switch /bazel. That's why neither the .sh nor the .bat ever ran. The Windows path now uses backslashes.
• aws-lc-sys build script. cl.exe failed with C1083 opening crypto\fipsmodule\rand/entropy/../../../../third_party/jitterentropy/jitterentropy-library/jitterentropy.h. MSVC joins relative includes without normalizing them, which made that path about 275 characters.
  • A crate.annotation patch changes the include to "jitterentropy.h". The aws-lc builder already puts that directory on the include path whenever jitter entropy is built, and the include sits inside #if !defined(DISABLE_CPU_JITTER_ENTROPY).
  • No other aws-lc relative include goes past 249 characters.
• Fixture paths. code-extraction's main and unit tests, along with 53 suites across 14 crates, read ../../tests/fixtures relative to the crate directory, which is where Cargo starts tests.
  • The launchers now cd into the crate's runfiles directory before running the test.
  • Every test now carries //tests:fixtures (160 files, 1.1 MB). Linux would have hit the same failures, but its run timed out before reaching those tests.
• Exit 9009 on five early tests. The probe, framing and private-fs tests failed because their .exe wasn't in the runfiles tree. The same launcher worked for later tests, and the cause isn't known yet. If it happens again, the Windows launcher now reports the missing binary and lists the runfiles tree.
• tracedecay-code-index and tracedecay-lsp: E0463 can't find crate for every dependency, with no extern-location note. The process_wrapper command line is 27,908 characters, so this isn't command-line truncation. A failure-only bazel aquery --include_param_files step now prints lsp's exact rustc arguments.

The Linux and macOS lanes of run 37511829834 are still running. They're the first check of the serde_json and unit-test feature fixes, so I'll dispatch db0abfdb once they finish rather than cancel them.

---

Generated by Claude Code

[ScriptedAlchemy]

Status at bcc51b4b (dispatched with run_os):

Linux at b7d27ef (job 112555776231). The whole suite ran in 3h15m: 151/160 pass, against 14 failures plus a timeout at 14b8ffd. The 9 failures, sorted:

Already fixed on head:

• search-eval:unit_test: 912022e8 re-pins the receipt.
• core_cli_suite: missing host-CLI example, fixed by TRACEDECAY_HOST_CLI_FIXTURE.
• code-extraction:unit_test: crate fixtures/, fixed by debb3f47.

Rustup reached through a redirected HOME ("missing manifest in toolchain 1.97.1", "syncing channel updates"):

• mcp:unit_test
• mcp_suite (affected-tests cases)
• runtime_acceptance_suite (compiler lookup, LSP negotiation)

The runner leaves CARGO_HOME/RUSTUP_HOME unset, so the launcher's $HOME fallback points into the test scratch dir. bcc51b4b exports both homes.

SDK journeys:

• transport_acceptance_suite: pnpm run build in sdks/typescript fails to spawn. bcc51b4b adds setup-pnpm to the test tools, as the Cargo-era test jobs had.

In-process vs nextest isolation:

• cli_non_interactive_test (core_cli_suite): "core publication was not held".
• session_store_read_cost: per-message writer work 1.3× against a 1.2× bound.
• session-temporal-store:unit_test: "full root candidate window: Unavailable" (to confirm).
• runtime_surface_acceptance: one MCP git-preview problem, not yet classified.

Build parity fixes, both host-wide:

• 1baabab6: rules_rust replaces a build script's CFLAGS/CXXFLAGS with the C toolchain's, so --action_env never reached cc-rs. Vendored C ran without -DNDEBUG everywhere and without /utf-8 on Windows, which is where the 347 Windows extraction Utf8Error/poisoned-LazyLock failures came from. The flags now travel as HOST_/TARGET_CFLAGS, which cc-rs appends.
• 3f51fa9e: .cargo/config.toml's LIBSQLITE3_FLAGS never reached Bazel; a crate annotation adds it.

Windows E0463. The params file does name every --extern. Locally, a missing, wrongly named, empty or directory extern each produces its own rustc message, so a bare E0463 comes from a rejection the locator doesn't record. f208a4e7 reruns one lsp compile with RUSTC_LOG=rustc_metadata::locator=info on failure.

---

Generated by Claude Code