[Patrick-Erichsen]

Solaris leaves unused space in sequential source packets, spending the shared scan deadline on extra model requests. Choose the fullest fitting packet among at most 16 open packets, using measured serialization costs, while preserving every original source unit, range, finding identity, and same-file chunk order. Every grouped prompt is checked against its actual UTF-8 byte size; an unexpected overflow restores the entire original singles plan in source order. Existing member/finding limits, overlapping-source separation, oversized singleton behavior, concurrency, workflow deadline, and model settings remain unchanged.

Retained-original source planning reduces requests from 632 to 552: developer 176→154, quality 178→156, discovery 172→152, meta 106→90, with approximately 1.3 seconds of planning work. Source contents/ranges/finding IDs and actual final packet limits are verified. These are reconstructed planner inputs; retained final meta messages differ from original pre-provider inputs, so this does not establish full-model completion or a user unblock. The red/green minimal regression reduces three requests to two.

Validation: 526 related producer controls; 349 combined native adoption controls with pinned shfmt v3.14.1; all 134 adopted source/dependency files match the final producer plus the reviewed predecessor. Independent producer and final adoption P0–P3 reviews completed clean after a regression proved global fallback preserves source ordering. The new suite runs in image builds.

Depends on #117.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.