[codex] Optimize profile sigma and dense radiance storage

[bout3fiddy]

Summary

• Fix profile-line reuse for retrieval pressure-state changes.
• Convert exact-route support-profile sigma storage from 64 B AoS rows to a wavelength-major []f64 total-sigma column.
• Split dense radiance prefetch storage into radiance-only []f64 plus optional Jacobian vectors so no-Jacobian gathers read only radiance data.
• Update retained cache/layout and spectrum tests for indexing, cache reuse, radiance-only gathers, and Jacobian-capable shape checks.

Validation

• zig build test-fast
• zig build test
• uv run scripts/linting/check-zig-styleguide.py ...
• uv run benchmark/run_benchmark_fast.py
• prek run --all-files

Fast benchmark residuals were unchanged versus the parent baseline; timing moved within canary noise on the forward paths.

---

Note

Medium Risk
Changes core session-cache invalidation and dense radiance/Jacobian memory layout on every forward and OE iteration; incorrect reuse or indexing would show up as subtle spectroscopy or Jacobian errors rather than obvious crashes.

Overview
Session cache and forward hot path: Profile-line reuse stamps no longer hash full LayerGrid support placement, so retrieval iterations that only change aerosol pressure can keep reusing spectroscopy rows instead of invalidating the cache. Support-profile line sigma is stored as a wavelength-major []f64 total column instead of 64-byte ProfileSupportLineValue rows, and dense radiance prefetch is split into []f64 radiance plus optional Jacobian vectors so no-Jacobian gathers avoid loading Jacobian-sized structs.

Python/Zig input boundary: The packaged O2 A reference scene is defined in Python (default_o2a_scene() / o2a_default.py); Zig drops built-in default scene rendering and the C exports zds_prepare_default_o2a and zds_default_o2a_input_json. Native JSON encoding maps inert NaN to null via native_json_value; route constants like stream count live in json.zig instead of a Zig defaults module.

Tooling and docs (secondary): ./verify is documented as canonical CI; pytest markers and dev deps are added; wheel smoke scripts point at test_wheel_install_smoke.py; cost-timing Zig target is renamed to cost-timing-forward-analysis. Retained benchmark and performance README numbers are refreshed.

^{Reviewed by Cursor Bugbot for commit f256914. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit f58874b. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f58874bc57

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1fa91c273f

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".