[jtravs]

Two Adams failure-semantics fixes, one commit each.

#129: a failed native Adams source no longer leaves the prediction in the state

• The bare std::swap(current_, adams_saved_) pairs in adams_fused_trial and adams_elementwise_trial become a SwappedAside scope guard (native/src/general/kernels.hpp), and adams_evaluate (coupled trial and startup) restores current_ before rethrowing.
• New test hook adams_march_fault("trial_source", …) throws once at the start of the next source evaluation, while the prediction is in the state.
• Regression test on the elementwise, fused and coupled paths: the checkpoint field and march snapshot equal those from a fault injected before the prediction (to 1e-12 relative). All three fail without the fix.

#651: record the completed rejected attempt before raising the floor

• NumPy reference (host/adams.py): _reject holds the floor error, and march_to raises it after _record, as _propose already does for accepted steps. Native (adams_march.cpp) does the same. A replay is unchanged (no row; the floor is raised first). Proposal, state, ledger and message are unchanged.
• On this floor path, a failure while releasing the attempt is kept as a note on the floor error, matching the reference (the TDSE abandoned-attempt test covers it).
• As the issue asks, the pre-refactor trace corpus is not regenerated: floor_after_rejections joins the existing CHANGED sets, and an explicit test checks that the new trace equals the corpus trace plus exactly one rejected row. The issue's witness scenario is tested on both marches.
• docs/dev/kernel-contract.md documents the floor ordering.

Checks: ruff, basedpyright, poe docstrings, clang-format, codespell; 252 targeted tests on CPU (Adams traces, native march traces, TDSE native, march lifecycle, session history, resident Adams, exponential march). Not run locally: clang-tidy, sanitizers, device builds (CI device-compile covers the builds; an NVIDIA A40 GPU test run will be reported here).

Closes #129
Closes #651

🤖 Generated with Claude Code

https://claude.ai/code/session_01XGENPwwkUUG9KSAwneHjgH

[read-the-docs-community]

Documentation build overview

📚 solprop | 🛠️ Build #35876910 | 📁 Comparing fbb154c against latest (a3ac8c4)

  🔍 Preview build  

2 files changed

± dev/kernel-contract.html
± _modules/solprop/backends/host/adams.html

[jtravs]

GPU lane on an NVIDIA A40 (CUDA 12.8.1, AMPERE86 build) at 8281643:

• native gpu tests: 31 passed, 1 skipped.
• solprop gpu tests: 289 passed, 1 skipped, 6 failed. The 6 are segfaults on the radial route on the device (test_monitor_reduction radial-* split removal, test_native_session_device radial/collar cases). They reproduce identically on main at a3ac8c4, so they predate this PR; tracked separately.
• The Adams march/trial modules (test_native_march_lifecycle, test_native_march_traces, test_resident_adams, test_adams_traces) pass against the device build: 138 passed (134 on main, plus this PR's 4).

The later commit 0bde86e only restructures host-side adams_march.cpp control flow to stay under clang-tidy's function-size limit; there is no device code change.

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-13T19:30:24.286125Z	fbb154c	New commits
🔒 Security Review	✅ Completed 2026-11-13T19:31:05.835293Z	fbb154c	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[jtravs]

Pre-existing radial device crash filed as #815.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!