evals(community): OsaurusAI/Laguna-XS-2.1-JANG_6M on Apple M5 Max

[jax-0n-git]

Crowdsourced model-compatibility run (see COMMUNITY_EVALS.md).

One contribution file; no other changes.

Environment: chip=Apple M5 Max · RAM 128GB · osVersion=26.6.2 · commit=cc19e77 · judge=self-judge · catalogHash=2e766b2b52f4cf37 · contributor=jax-0n-git

[jjang-ai]

Cross-family audit: this is useful historical Laguna data (AgentLoop 117/139, loop-free 89.93%, 44.60 tok/s, 115 ms TTFT, 26,356 MB peak), but commit cc19e77 predates the global tool-kill-switch deletion (#2505), bounded partial-read steer (#2528), and exit telemetry (#2520). The aggregate does not retain exit origins or cache counters, so the 22 scored AgentLoop misses cannot yet be attributed to shared loop logic versus Laguna parser/reasoning/cache behavior. PR #2546 adds privacy-safe exit/recovery/KV-SSM-L2-MTP totals. Required follow-up is a same-catalog current-head rerun plus visible Release-app tool-result continuation and Stop/restart proof; this report alone is not a current regression verdict.

[jjang-ai]

Current-main diagnostic rerun (not a merge/release gate): Osaurus 55e65b4, pinned vMLX bf8b31995195fffd833968658f14c707317eaa70, model dealignai/Laguna-XS-2.1-JANG_6M-CRACK, MTP explicitly off.

Selected current AgentLoop smoke: 6/6 passed (direct answer, Stop/cancel after first tool, folder listing, repeated-call nudge, targeted read, file write). Objective tool/file checks passed; the rubric-backed direct-answer row used self-judge because no external judge was configured, so this is not a replacement full-suite score.

Measured:

• final-text decode: mean 36.8 tok/s (35.7–38.5, n=4)
• TTFT mean 1,185 ms (121–4,632)
• peak phys_footprint 21,006 MB
• disk L2 delta +11 hits / +133 misses / +46 stores
• cancel row exited cancelled with no later write, but first action took 69,552 ms
• all tool-only steps still report unavailable_tool_call_before_vmlx_info, so their tok/s is missing

Important host caveat: a separate user-owned vmlx-serve process was discovered after the row, and host swap rose from 0.36 GB to 24.8 GB. Functional/tool assertions are retained as diagnostic evidence, but perf/footprint are contaminated and not comparable to #2492's historical report. No other family run will start until that process is gone and the preflight checks all vMLX/MLX processes, not just Osaurus.

Artifacts on proof Mac:

• /private/tmp/osaurus-current-proof/laguna-xs-6m-smoke.json SHA-256 6c19b95d65cac66f484604c6c1cecaa3d153114fc616eabd44bb969bea2a90c5
• log SHA-256 d51c0fc746b6a1562a978b930a86a67779a958420b593be11c63944595d056fb

Status: PARTIAL. Clean rerun + exact-head Release UI proof remain required.