Gap analysis: hand-edited Spark state with no modeled authority

[gunbai-bot]

The operator asked for an accumulation of what has been hand-edited on the live Sparks while the fleet is assembled by hand.

The reason it is worth a document rather than a plan line is one finding: a convergence run today would degrade the fleet without failing. The modeled desired unit differs from the live unit on BOTH serving hosts in three directions at once, all pointing the same way:

• context 131072 modeled against 1048576 live -- below the 400k floor gunbc.model.choice exists to defend
• served model gpt-oss modeled against hf.co/antirez/deepseek-v4-gguf live
• OLLAMA_NUM_PARALLEL appears nowhere in the corpus, so convergence does not overwrite the slot count, it DELETES it and the runner falls back to a default

Nothing on that path refuses. The unit rewrites, the service restarts, it serves -- narrower and serialized.

Everything was read off the live hosts and compared against the module named as its authority. Desired side: gunbc.spark.serving_unit_render spark_serving_desired_user_unit_text. Live side: the unit file and /api/ps per host.

Also rostered: the second pair (spark-c2b1, spark-ac79) is serving while unknown to cell_role, the fleet slot roster, and network_identity_subsumption; residency is unmodeled on both sides; credentials were hand-carried past gunbc.spark.credential_workflow .

One row turns back on my own #9897: it keys realization identity on argv, and every runner here is bare ollama serve with its configuration in Environment= lines -- so on this fleet that identity discriminates almost nothing. Widening it needs a ruling on which variables are configuration and which are ambient.

Doc only -- no code, no substrate.

🤖 Generated with Claude Code

https://claude.ai/code/session_01U397y4s3dSBof7vGPAX87G

[gunbai-bot]

CI on this branch is red for a reason that is not in this diff, and I want that on the record rather than silently retried.

This PR is one new markdown file, +145 lines, nothing else (git diff --stat origin/main...HEAD shows a single path under docs/plans/). It touches no .dag, no Rust, and no generated artifact.

The failing lane is required-witnesses-build, and its failure is:

required-ci: regen FAIL generated surface drift: gunbc_rust_source_type_bindings.rs

That same failure is present on main itself at d61f8c9bc, which is the commit this branch merged. Run 33584275111 on main fails with the identical line. The drifting artifact is untouched by this branch — git diff origin/main...HEAD -- gunbc_rust_source_type_bindings.rs is empty.

So this is inherited, not caused. I am deliberately not regenerating the artifact on a docs-only branch: that would put a generated-surface change in a PR whose entire subject is a planning document, and the regeneration belongs wherever the drift was introduced. Main has since advanced past d61f8c9bc (currently 3138c4617); once a green main is available I will merge it and let the lane re-run.

If the drift turns out to persist on current main, that is a main-side repair and I will raise it separately rather than smuggle it through here.

— sent from eager-pike-541

[briansrls]

Operational supersession: #10001 is now the controlling immediate path to safe Spark fleet convergence.

Before this document can govern operations, its stale PR body and repeated paragraph must be corrected: context and OLLAMA_NUM_PARALLEL are not current divergences; the live blockers are the exact desired serving realization, safe treatment of an active stale process, and the current srv6 role statement. The operator has now withdrawn the dedicated-training assignment for srv6 for this phase.

Do not broaden #9991 into implementation. Link the corrected gap record to #10001, where completion requires one exact-host wet plan→apply→full readback→zero replan→terminal health transaction plus a second zero-effect noop.