[jonyoder]

A depender version that forms a run of one (no identical neighbour) now covers the gap up to the next published version, so learned conflicts of adjacent turned-down versions merge into one span instead of one span each. Fixes two real PPM out-of-memory solves. For rstudio/package-manager#22613.

---

Why it is sound

• A run of one for v is now [v, next published version).
• No published version lies strictly inside that gap, so the solver can never pick anything the old span [v] excluded.
• Pins and failures are unchanged; only failure text can get shorter (a range where it listed versions one by one).

Measurements

Measured in PPM (f8679e25ba) with base = v0.21.0 and new = this change on v0.21.0 (go-pubgrub v0.5.0), before the merge with v0.22.0/v0.23.0. Production cache sizes, 6 GB container, 8 CPUs, machine under load. RS = RunningSatisfier.

Case (RSF)	base RS on	base RS off	new RS on	new RS off
s3fs + anndata, 30 envs (11-12)	OOM at 7 s (4.9 GB heap)	hit 5.5 GB heap cap at 929 s	13.4 s, 1.35 GB, 59 pins	115.9 s, 1.60 GB, 59 pins
s3fs + scanpy, 30 envs (11-12)	OOM at 8 s (5.1 GB)	not done at 600 s	24.8 s, 1.44 GB, 94 pins	223.9 s, 1.87 GB, 94 pins
f1 (11-12, parity run)	OOM at 11 GB		184 s, 3.3 GB

The anndata pins equal uv's exactly. On the 09-27 RSF there is no gain on f1 (2.3 -> 2.3 s), s1 (noise) or s2-105 (both builds hit the 5-minute deadline), and a small gain on s3-400 (RS on 30.9 -> 27.1 s, off 20.5 -> 16.9 s). Same pins, failures and warnings in every pair; heap peaks within noise. Rounds and Dependencies calls do not drop: the gain is smaller learned clauses.

scanpy: uv pins 95, PPM 94. On Python 3.11 PPM picks anndata 0.12.6 where uv picks 0.12.19 + scverse-misc 0.0.3. s3fs+scanpy+anndata==0.12.6 resolves in uv, so PPM's answer is valid but older; it comes from PPM's curated layer, not this change (base never finishes this case).

Parity

114 + 111 environment sets on the 09-27 and 11-12 RSFs: no pin differences. Every differing cell is failure text only (shorter explanations), in 6 sets per run. Some cap reruns (sets that hit the 4.5 GB harness cap in one build only) were not finished.

Checks

• go test -race ./... and golangci-lint (0 issues) pass after merging origin/main.
• RoboRev 5255 (code + CHANGELOG) and 5256 (merge) passed with no findings.
• Independence audit (go-pubgrub audit.sh, scoped to this worktree): PASS, no known reference retrieved in 236 sessions. Process finding only; it says nothing about model weights. Two cross-repo notes are downloads of the uv release binary, not source.

🤖 Generated with Claude Code