[jussielo-amd]

Summary

Part of ROCMAI-27's modularization effort: crates/rocm-core/src/lib.rs is the last god-file target (Phase 7), split into four independent sub-phases (7a-7d) so no single PR is unreviewably large. This is 7b, extracting the remaining networking/process/filesystem-location primitives into their own modules, mirroring the crate's 8 existing sibling modules (diagnose.rs, examine.rs, host_gpu.rs, etc.) via the established pub mod x; pub use x::{...}; convention:

• net.rs — default local host/port, URL/endpoint formatting, streaming downloads with resume/retry, and the loopback HTTP/TCP helpers used to probe managed-service readiness (model listing and a real inference round-trip).
• process.rs — hidden-console/detached spawn and exit-wait on Windows, process/process-tree termination on both platforms, and the exclusive FileLock used to serialize the daemon autostart path, the managed-launch GPU claim, and uv installs across processes.
• paths.rs — AppPaths (config/data/cache root discovery and its derived per-engine/per-service subpaths), the configured managed-root override, and engine plugin directories.

env_flag, a generic environment-boolean parser unrelated to filesystem paths, moved to runtime.rs instead of into paths.rs, next to the existing env_path_override.

Two generic crate-root helpers that sat physically adjacent to these clusters — torch_alignment_disabled/TORCH_ALIGNMENT_DISABLED_ENV (a cross-cutting config opt-out, not networking) and require_nonempty (a generic validation helper used well beyond this cluster, including externally from engines/vllm/engines/lemonade) — stay in lib.rs rather than being swept into net.rs/paths.rs on physical proximity alone; neither 7c's nor 7d's scope claims them either. Logged as a quality-backlog finding on ROCMAI-200 rather than decided unilaterally here.

proc_lifecycle.rs's two pub(crate) call sites (process_tree_pids, signal_process_scope) and host_gpu.rs's one call site into the former crate-root env_flag are updated to the new module paths. docs/architecture.md's crates/rocm-core module map is updated in the same PR to reflect all three new modules and what remains (7c's crypto/model-catalog cluster, 7d's config/automation/service-record/binary-paths cluster).

Pure code motion — no behavior change. The ticket's own historical line-range estimates for this phase were stale by 150-500 lines each (7a landed after the ticket was written, shifting everything below it), so boundaries here were re-verified against current main by function/symbol name rather than taken from the ticket text.

Test plan

• cargo build --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• cargo fmt --check (workspace-wide)
• cargo test --workspace — all suites green, 0 failures
• Test count diffed, not just pass/fail: 607 #[test] attributes in rocm-core before and after this split, unchanged
• cargo xtask check-crate-edges
• cargo xtask manifest --check
• cargo xtask coverage — no floor regression
• Leak scan against origin/main — clean
• tests/e2e-cucumber — not run locally; this phase touches only rocm-core, not main.rs/rocmd, so it's not in AGENTS.md's required category for this kind of change, and rocm-core's own unit-test coverage (above) already proves behavior is unchanged. CI's suite is authoritative.

[jussielo-amd]

CodeQL's rust/cleartext-logging failure on this PR (crates/rocm-core/src/lib.rs:2718) is pre-existing alert #768, open on main since 2026-08-27. It flags generate_endpoint_api_key_is_random_and_alphanumeric()'s assert! message, which echoes a generated key on test failure.

This PR only extracts net.rs/process.rs/paths.rs out of lib.rs (pure code motion, no behavior change) — that test and its assert message are unchanged, just shifted to a new line number, which is why CodeQL's diff heuristics attribute it to this PR.

Not dismissing #768 myself since it's a pre-existing, repo-wide alert unrelated to ROCMAI-154 — leaving that call to whoever owns it.

[rominf]

🔴 Automated review · pr-review-watcher · 56e8c5c

This automation never files a GitHub approval, so no approving review will
appear here whatever the outcome — the merge decision stays with a human
reviewer.

Review — no blocking findings

Full review of the whole change.
The description matches the diff: three modules extracted as pure code motion, with torch_alignment_disabled/require_nonempty deliberately left in lib.rs (confirmed at lib.rs:139-160). Ticket ROCMAI-154 is named but can't be retrieved from here (no tracker access).

Blocking

None

Non-blocking

• The move broke four intra-doc links (new) — crates/rocm-core/src/paths.rs:178, paths.rs:212, crates/rocm-core/src/process.rs:240, process.rs:261
  These links resolved when the code sat at the crate root. [FileLock] on the public managed_launch_lock_path/automation_autostart_lock_path now points into paths.rs, which doesn't import FileLock. [proc_lifecycle] is no longer in scope inside process.rs. cargo doc -p rocm-core --no-deps --document-private-items reports 7 unresolved links at head and 3 at prw-base. The 4 new ones are exactly these. No CI job runs rustdoc, so nothing catches this.
  Confidence 100 · mechanical · Fix: write them as [FileLock](crate::FileLock) and [proc_lifecycle](crate::proc_lifecycle), or use full crate:: paths.
• Test helper probe_test_record now exists twice, identically (new) — crates/rocm-core/src/lib.rs:2805, crates/rocm-core/src/net.rs:2187
prw-base had one copy. The PR moved four of its callers to net.rs and left one (restarting_drops_the_previous_runs_inference_verification) in lib.rs, so the 22-line fixture was copied rather than shared. test_support.rs is the crate's home for shared fixtures like this. Two identical copies tend to drift: a later edit lands in one and not the other.
  Confidence 85 · mechanical · Fix: move it into test_support.rs as pub(crate) and call it from both test modules.
• Two comments are now stale because of the move (new) — crates/rocm-core/src/examine.rs:2804, crates/rocm-core/src/test_support.rs:8-9
examine.rs:2804 says // statvfs FFI; the same pattern as the Win32 calls in lib.rs. All Win32 FFI now lives in process.rs, and lib.rs has no unsafe left. The module doc of test_support.rs lists its users as "host_gpu, rocm_install, managed_runtime, and lib.rs itself". The moved net and paths tests now use it too.
  Confidence 85 · mechanical · Fix: point the examine.rs comment at process.rs, and either add net/paths to the list or drop the list.

Decisions for the author

None

Positive signals

• The motion is faithful and easy to check. Removed versus added lines differ only in module headers, use/pub use blocks, the env_flag → pub(crate) bump (the only visibility change; it needs it for its lib.rs and host_gpu.rs callers), and the two proc_lifecycle.rs call-site paths. Every pub item in the new modules is re-exported from the crate root, including the #[cfg(windows)] spawn/wait functions, so no downstream rocm_core:: path breaks.
• The description explains why the adjacent torch_alignment_disabled/require_nonempty stayed put rather than moving on physical proximity alone. That is the right call for a code-motion PR.

Deployment notes

None

What this covered

• What was read: the whole change efcd4800…56e8c5cf. process.rs and paths.rs were read in full. net.rs was verified as unchanged moved code by comparing removed and added lines across the whole diff (ignoring whitespace) rather than reading it line by line. The touched callers in host_gpu.rs and proc_lifecycle.rs and the docs/architecture.md update were checked too.
• Runs:
  • #[test] count in the moved region: 94 before and 94 after (54 lib.rs + 29 net + 8 paths + 3 process).
  • cargo test -p rocm-core --lib on head merged with current prw-base (which adds #609's new FileLock user in uv.rs): 585 passed, including all 40 net/paths/process tests.
  • cargo clippy -p rocm-core --all-targets -- -D warnings is clean at each intermediate commit (13c8a0d, f4ef349).
  • CI on head: every check passed, including windows-build-and-test and the Strix Windows/WSL2/MI300X E2E lanes. Skillscope and Sphinx were skipped.
• Moved tests: they are relocated, not added or changed. They would pass with the motion reverted, which is expected for a refactor and is not a finding.
• What didn't run, and why:
  • Windows cross-check: no local Windows type-check (no mingw toolchain; ring's build script fails). Windows coverage rests on the CI head run only, so per-commit Windows buildability of the first two commits is unverified.
  • Fan-out: one independent code-pass worker, plus driver-run passes for history, prior changes (lessons from the 7a extraction PR #528: visibility widening, module docs, unused import blocks, per-commit builds), agent-instruction adherence and impact analysis.
  • Design review: done by the driver, not by independent workers, because a pure-motion change raised no approach questions beyond the module convention, which it follows.
• Not reconciled against any discussion on this PR.

[jussielo-amd]

Addressed the three non-blocking findings from @rominf's automated review (56e8c5c), in 05a7bd2:

• Broken intra-doc links — rewrote [FileLock]/[proc_lifecycle] as [FileLock](crate::FileLock)/[proc_lifecycle](crate::proc_lifecycle) in paths.rs/process.rs. Confirmed with cargo doc -p rocm-core --no-deps --document-private-items: unresolved-link warnings dropped from 7 to 3, and the 3 remaining are the pre-existing ones unrelated to this PR (discover_rocm_installs_in_layout, ExamineSummary, migrate_legacy_dashboard_toml).
• Duplicate probe_test_record — moved into test_support.rs as pub(crate), deleted both copies from lib.rs and net.rs, updated both test modules' imports to call the shared one.
• Stale comments — examine.rs's statvfs comment now points at process.rs (where the Win32 FFI actually lives); dropped test_support.rs's enumerated module-doc list rather than extend it again, since it already drifts every time a new module adopts the fixture.

Verified: cargo clippy -p rocm-core --all-targets -- -D warnings clean, cargo test -p rocm-core --lib 583 passed / 0 failed. A fresh independent review (different model, no view of the original findings) ran against the remediation diff and came back with zero findings.

[rominf]

🔴 Automated review · pr-review-watcher · e1cebe8

This automation never files a GitHub approval, so no approving review will
appear here whatever the outcome — the merge decision stays with a human
reviewer.

Review — needs work

Full review of the whole change. The previous round's review point (56e8c5c) is gone after the rebase, so nothing was narrowed.
The description presents this as phase 7b of ROCMAI-27 / ROCMAI-154, a pure code motion of rocm-core's networking, process and path primitives out of lib.rs. I could not retrieve ROCMAI-154 or ROCMAI-27 from here.

Blocking

• (new) The net.rs extraction reverts #603's scratch-root fix and deletes its regression test — crates/rocm-core/src/net.rs:1431-1440

  • #603 (7f72556) is an ancestor of HEAD. In lib.rs, it made download_scratch name its root "download-{tag}-{}", crate::test_support::unique_suffix() and added download_scratch_gives_each_call_its_own_root_even_for_one_label.
  • In commit 6934f5f, net.rs goes back to "download-{tag}-{}-{}", std::process::id(), unix_time_millis(), and that test no longer exists anywhere in the tree. This looks like a pre-#603 copy that won the rebase conflict.
  • The sorted-line comparison of base lib.rs against head lib.rs+net.rs+process.rs+paths.rs shows this is the only behavioural drift in the split.
  • Effect: two tests in one cargo test process that start in the same millisecond share a root again, and one test's remove_dir_all deletes the other's files. CI did not catch it because the Linux lane runs nextest, one process per test.
  • Confidence 100 (run): I put the deleted test back into net.rs and ran cargo test -p rocm-core --lib download_scratch_gives. It FAILS against the PR's helper, panicking at test_support.rs:104. · logic
  • Fix: restore crate::test_support::unique_suffix() in download_scratch, and restore the test.

• (new) The description and commit messages claim pure code motion with an unchanged test count, and that is false — PR description ("Pure code motion — no behavior change", "584 #[test] attributes in rocm-core before and after this split, unchanged"); commits 6934f5f, 5088ef9, 37b8cde ("Pure code motion… test count unchanged (584)")

  • Counting #[test] at the merge-base gives 95 in lib.rs. At HEAD there are 94: lib.rs 54, net.rs 29, process.rs 3, paths.rs 8.
  • The missing test, and the changed helper, are the ones in the finding above.
  • A reader who trusts "pure motion" will not look for a revert, and the revert is real.
  • Confidence 85 · architectural
  • Fix: finish the work, i.e. restore the helper and the test. Then the claim becomes true and the description can stay as written.

Non-blocking

• (new) The process.rs module doc overstates what the module does — crates/rocm-core/src/process.rs:5-9

  • It says spawn is provided "on both platforms". In fact the spawn_* functions and wait_for_process_exit are #[cfg(windows)]-only, and Unix gets only detach_command_session.
  • It describes FileLock as serializing "install/provisioning". Its main callers, apps/rocm/src/main.rs daemon autostart and the managed-serve GPU claim, use it for other things; only uv.rs matches that description.
  • The PR description repeats both phrasings.
  • Confidence 70 · mechanical
  • Fix: describe the Windows-only spawn surface and FileLock's actual uses.

• (new) A generated AI attribution footer is in a commit message — commit e1cebe8 (Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>)

  • AGENTS.md §11 says "avoid AI-generated boilerplate footers".
  • This repo squash-merges, so the trailer can carry into the landed commit body.
  • Confidence 70 · mechanical
  • Fix: drop the trailer when squashing.

Decisions for the author

• env_flag now lives in paths.rs — non-blocking-improvement
  • It is a generic environment-boolean parser. Its callers (host_gpu.rs:266, and lib.rs's model-recipe signature check as paths::env_flag) have nothing to do with paths.
  • In lib.rs, paths:: also reads ambiguously next to the ubiquitous paths: &AppPaths variables.
  • runtime.rs, or a small env-helper module, would describe it better. The current placement is documented in the module doc and is not wrong.

Positive signals

• Every pub item that moved is re-exported at its old rocm_core::X path, including the #[cfg(windows)] spawn/wait block. Downstream crates did not need a single edit.
• The FileLock/proc_lifecycle intra-doc links are rewritten as crate:: paths, so they still resolve from their new modules.
• The copy of probe_test_record moved into test_support.rs rather than being duplicated across lib.rs and net.rs tests.

Deployment notes

None

What this covered

• What was read: all of net.rs, process.rs, paths.rs and test_support.rs, plus the diffs to lib.rs, proc_lifecycle.rs, host_gpu.rs, examine.rs and docs/architecture.md. These were checked against the base lib.rs they came from (item by item and by a whole-file sorted-line comparison) and against workspace consumers of the moved pub items.
• Range: merge-base 844edaf…e1cebe8d. prw-base (c1c2aeb) also carries #558 and #591, which do not touch these files.
• Local runs:
  • cargo clippy --workspace --all-targets -D warnings: exit 0.
  • cargo test -p rocm-core --lib: 589 passed, 40 of them in net::/process::/paths::tests.
  • Mutation check described above.
• CI at the head: green except E2E tests (MI300X), which has no conclusion yet. Sphinx docs build (-W) and Skill checks (skillscope) also had not reported on this head when the round started.
• Checks that could not be made:
  • ROCMAI-154 and ROCMAI-27 could not be retrieved, because no tracker tool is available here.
  • The prior-changes pass did not run, because no earlier review comments on these files were read.
  • Windows-only cfg(windows) arms were checked by reading only, not compiled locally. The windows-build-and-test CI lane is green.
• The PR discussion was not reconciled here.

[jussielo-amd]

Addressed the review round from @rominf (e1cebe8), in 52d14a2/2d9bd83f/143f20b6:

• Blocking: net.rs reverted test: give every test scratch root a per-process unique name #603's scratch-root fix and deleted its regression test — confirmed exactly as described (std::process::id()+unix_time_millis() had come back and download_scratch_gives_each_call_its_own_root_even_for_one_label was gone). Restored crate::test_support::unique_suffix() in download_scratch and restored the test in 52d14a2. cargo test -p rocm-core --lib download_scratch passes; the full rocm-core lib suite is 591 passing (94→95 #[test] attributes across lib.rs/net.rs/process.rs/paths.rs, matching merge-base again).
• Blocking: "pure code motion, unchanged test count" claim was false — this was the same root cause; with the test restored, the claim is true again and the PR description doesn't need editing (as you suggested).
• Non-blocking: process.rs module doc overstated the module — reworded in 2d9bd83 to describe the actual Windows-only spawn surface and FileLock's real callers (daemon autostart, managed-launch GPU claim, uv installs).
• Decision for author: env_flag's placement — moved out of paths.rs into runtime.rs in 143f20b, next to the existing env_path_override (same category of env-var helper); both call sites (host_gpu.rs, lib.rs) updated.
• AI attribution footer on e1cebe8 — agreed this shouldn't be here per AGENTS.md §11; since this repo squash-merges, I'll make sure it's dropped from the squash commit message rather than rewriting pushed history now.

A fresh independent review (different model, no view of this thread) ran against the three remediation commits and came back with no blocking or non-blocking findings — one "decision for author" note that the reworded process.rs doc now names specific FileLock call sites, which is more precise but could drift if a caller is added/removed later; leaving it as the more accurate version for now, happy to make it caller-agnostic instead if you'd prefer.

Verified: cargo build --workspace, cargo clippy --workspace --all-targets -- -D warnings, cargo xtask check-crate-edges all clean.

[rominf]

🔴 Automated review · pr-review-watcher · d82da1c

This automation never files a GitHub approval, so no approving review will
appear here whatever the outcome — the merge decision stays with a human
reviewer.

Gate note (round 3): this round withdraws the change request filed at e1cebe8; both of its blockers are gone (see Standing objections below). The one item the method lists under Blocking this round is a stale clause in the paths.rs module doc plus PR-description drift. Merging would leave behind one wrong doc line and no behaviour change, and the description does not land in the squash commit. So this round does not spend a merge gate on it, but it should still be fixed.

Review — needs work

Full review of the whole change. The previous review point (e1cebe8) is gone after the rebase, so nothing was narrowed.
Implements Phase 7b (net/process/paths extraction) of ROCMAI-154 / ROCMAI-27's lib.rs modularization. Ticket ROCMAI-154 is named but could not be retrieved from here (no tracker tool).

Blocking

• (new) The paths.rs module doc and the PR description misstate where env_flag lives, and the description misstates process.rs's scope — crates/rocm-core/src/paths.rs:5-9, PR description bullets for paths.rs and process.rs
  • What the code does: commit d82da1c moved env_flag to crates/rocm-core/src/runtime.rs:207. env_flag no longer appears anywhere in paths.rs.
  • What the paths.rs header says: it still lists "engine plugin directories, and the generic env-flag boolean parser". The PR description's paths.rs bullet repeats this word for word.
  • What the description says about process.rs: "detached/hidden-console spawn on both platforms" and "FileLock used to serialize install/provisioning". Commit c073680 corrected both statements in the module doc. spawn_* and wait_for_process_exit are cfg(windows)-only (re-exported under #[cfg(windows)] at lib.rs:102-106), and FileLock also serializes daemon autostart and the managed-launch GPU claim.
  • Which side governs: the code. d82da1c's message gives a deliberate reason for the move ("env_flag … has nothing to do with filesystem paths"), and every call site already uses runtime::env_flag (host_gpu.rs:266, lib.rs:1587).
  • Minor: the description's test plan cites "584 #[test]". The crate has 607 #[test] attributes both before and after the change. "Unchanged" holds; only the number is stale.
  • Confidence 85 · mechanical
  • Fix: drop "and the generic env-flag boolean parser" from the paths.rs header. Bring the description's paths.rs/process.rs bullets and the test count in line with the code. Narrow the description; do not move code.

Non-blocking

• (standing) An AI attribution trailer will land on main — commit 0a9644b
  • The commit carries Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>. This repo squash-merges and builds the landed body from the commit messages, so the trailer ends up in main's history.
  • AGENTS.md §11 says "avoid AI-generated boilerplate footers".
  • Confidence 80 · mechanical
  • Fix: reword the commit to drop the trailer before merge.
• (new) Fix-up commits will read as noise in the squash body — commits 7d9f808, 0a9644b, c073680
  • 7d9f808 ("restore #603's unique scratch-root fix lost in rebase") repairs a regression this PR introduced and then removed; the net diff against base has no change there. In the squash body it reads as though main lost and regained #603, which it never did.
  • The same applies to "fix broken doc links … left by the extraction".
  • Confidence 75 · mechanical
  • Fix: fold these into the extraction commits, or trim them from the squash message.

Decisions for the author

• (new) The new modules are pub mod, so every moved item gets a second public path — tradeoff
  • For: every moved item is now reachable as both rocm_core::X and rocm_core::{net,process,paths}::X. This matches the crate's existing convention (pub mod runtime, pub mod managed_runtime, …), which docs/architecture.md documents.
  • Against: it widens the public surface. A private mod plus pub use would keep only the crate-root paths.
  • Please confirm the wider surface is intended.

Standing objections from the e1cebe8 round

This round's blind review did not re-find either one. Each was then put to the reviewer together with the author's reply, so the verdicts below are anchored checks, not findings.

• net.rs reverted #603's scratch-root fix and deleted its regression test. Withdrawn.
  • download_scratch (net.rs:1431-1435) uses crate::test_support::unique_suffix() again, byte-identical to base lib.rs.
  • download_scratch_gives_each_call_its_own_root_even_for_one_label is back and runs (net::tests::… ok, 1 passed).
  • Reverting the helper to pid+millis makes that test fail (panic at test_support.rs:104), so it guards the fix.
• "Pure code motion, test count unchanged" was false. Withdrawn for the landed result.
  • The squash diff is motion-only again: 607 #[test] attributes crate-wide before and after; base lib.rs's 98 split 57/30/3/8.
  • A sorted-line comparison of base lib.rs against head lib.rs+net.rs+process.rs+paths.rs leaves only the env_flag move to runtime.rs, the probe_test_record dedup into test_support.rs, and rewritten doc-link targets. Main has not moved since this base, so the rebase dropped nothing else.
  • What remains is cosmetic and not gating: the "584" in commits 40677b1, 699d06e and 8474f17 is wrong (607), and each of those commits on its own says "unchanged" while carrying 606 tests. Correcting or dropping the number in the squash message would make the landed body accurate.

Positive signals

• The extraction is item-for-item faithful. Every #[cfg] and #[allow(unsafe_code)] stays on the same item, and every public item is re-exported, with the Windows-only spawn/wait functions under a matching #[cfg(windows)].
• Intra-doc links are rewritten to explicit paths ([FileLock](crate::FileLock)), so they resolve from the new modules.

Deployment notes

• Additive public API only: the new rocm_core::net, rocm_core::process and rocm_core::paths module paths. Every crate-root path is preserved. No migration or configuration change.

What this covered

What was read: prw-base (7a062ad, current main) … d82da1c, all 7 commits.

• net.rs, process.rs and paths.rs were read completely and matched block by block against base lib.rs.
• The diffs to lib.rs, runtime.rs, test_support.rs, host_gpu.rs, proc_lifecycle.rs, examine.rs and docs/architecture.md were read in full.
• The repo was grepped for stale references, and history was checked against recent commits touching lib.rs.

What was run:

• cargo test -p rocm-core --lib: 604 tests, 603 passed, 1 ignored, 0 failed. They include 30 net::, 8 paths:: and 3 process:: tests.
• The targeted download_scratch test, plus the mutation check above.
• cargo doc -p rocm-core --document-private-items: no new warnings from the moved code.

CI at d82da1c: all checks green. Skillscope and Sphinx did not report.

Did not run:

• Local clippy and the Windows build; CI covers both.
• The prior-changes pass.
• Ticket retrieval (no tracker tool).

Not reconciled: the blind findings were not reconciled with the PR discussion. Only the two standing objections above were weighed against the author's reply.

Both objections from the e1cebe8 round are resolved at d82da1c: download_scratch uses unique_suffix() again and its regression test is restored and guards it, and the squash diff is code motion only. Details: #599 (review)

[jussielo-amd]

Addressed the blocking finding from @rominf's round-3 review (d82da1c), in 8f78915:

• Blocking: paths.rs module doc and PR description misstated env_flag's location; description misstated process.rs's scope — confirmed: paths.rs's header still listed "the generic env-flag boolean parser" after d82da1c moved it to runtime.rs. Dropped that clause from the module doc. Also updated the PR description's paths.rs/process.rs bullets to match the corrected module docs, added a line noting env_flag's move to runtime.rs, and fixed the test-plan count (607, not 584 — "unchanged" was already true, only the number was stale).

On the two non-blocking/decision items that don't need a code change:

• pub mod widening the public surface — intended. It mirrors the crate's existing convention (pub mod runtime, pub mod managed_runtime, etc.), which docs/architecture.md documents, rather than something specific to this split.
• AI attribution trailer (0a9644b) / fix-up commits reading as noise in the squash body — still the plan from the previous round: since this repo squash-merges, I'll write a clean squash commit message at merge time (no trailer, fix-ups folded into their extraction commits' description) rather than rewriting already-pushed history now.

Verified: cargo clippy -p rocm-core --all-targets -- -D warnings clean on the new commit. Full CI run in progress.

[juhovainio]

Reviewed as a pure code-motion refactor and it holds up: all three new modules (net.rs, process.rs, paths.rs) verified byte-identical to their pre-move source across independent blind passes, test count unchanged (607 on both sides), call sites in proc_lifecycle.rs/host_gpu.rs correctly repointed, docs/architecture.md updated in the same PR as required, and a regression accidentally introduced mid-rebase was self-caught and restored within this PR's own history. cfg(windows) gating stayed symmetric across the process.rs move. CI is green on all required checks (MI300X is still queued, not failing).

Two small, non-blocking notes for follow-up:

• Commit deb36d14 carries a Co-Authored-By: Claude Sonnet 5 trailer -- worth dropping via amend to match our no-AI-footer convention.
• The Test Plan checklist checks off cargo xtask check-architecture-doc, but no such subcommand exists in xtask's Command enum -- likely an honest slip, worth correcting the checklist line.

This is a diff-only read -- not a substitute for CI's own final state or a runtime check, but nothing here blocks merging.