feat(bernese): USER.CPU maxjobs sizing + V_CLUFIN/V_CLU plumbing (RH-006)

[alfieprojectsdev]

RH-006 (plumbing) — USER.CPU maxjobs sizing + V_CLUFIN/V_CLU (gap #13)

Stacked on #38; base retargets to main once #38 merges. Ships the orchestrator plumbing for the biggest R740 lever; the empirical clustering value is an R740 tuning task (see below).

Problem

Every daily run spent ~40 min of ~2 h in PID 502 GPSCLU_P — a single-core GPSEST at 99% CPU. Cause: V_CLUFIN=A auto-clustering collapsed the whole network into one dense normal-equation inversion, unparallelized. On the R740 this is the single highest-value tuning win; without it, 270 stations/day is infeasible.

Change (plumbing)

• cpu_config.compute_maxjobs(physical_cores, ram_gb=…, reserve_cores=…) (task L) — maxjobs from physical cores, not logical CPUs (sub-solves are FPU-bound; hyperthreads sharing an FPU don't help), RAM-capped (floor(ram_gb / ram_per_job_gb)) and reserve-aware.
• cpu_config.set_user_cpu_maxjobs(text, n) — rewrites the localhost maxjobs field in USER.CPU, leaving command/speed/other CPU rows/MSG lines intact.
• PCFContext now exposes v_clu (default 10) and v_clufin (default "A"); the template templates both — V_CLUFIN was absent from the template entirely.

Correction to the readiness doc

The real PCF shows V_CLUFIN is a mode flag (A auto / N skip), not a cluster-size number as the readiness doc implied. A produced the one-giant-cluster solve. The orchestrator can now inject a chosen V_CLUFIN + a core-sized maxjobs, but the value that actually parallelizes the final solve is empirical and needs the R740's core count + real timing (BRN-001, not installed). So RH-006 stays PARTIAL — mechanisms shipped, tuning value deferred.

Tests

test_cpu_config.py +13 (maxjobs cores/reserve/floor/RAM-cap/validation; set_user_cpu_maxjobs localhost-only rewrite + no-op + reject; PCFContext clustering defaults/override; template exposes the vars). 88 pass, ruff + mypy clean. Verified via the rh-004 worktree venv (no uv sync — BPE running). Worktree .trees/rh-006-clustering.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 100bfef5-906d-49a3-b95b-167cd93ef590

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/rh-006-clustering

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}