Branch-comparison and infrastructure assessment of the CatherineThomas-NOAA/global-workflow dev_gsiupd2 line (the pre-coupled-modeling DA development trajectory) against the current forked develop baseline. Captured to anchor the upcoming SDD work plan to fold v17 coverage into the EIB MCP / GraphRAG knowledge base.

dev_gsiupd2 is the current evolutionary tip of the v17 DA line. It is not a sideways fork — it is 21 commits straight ahead of develop — but its content culminates a divergent trajectory that pre-dates the upcoming coupled-modeling paradigm shift. Catherine Thomas's team drives experiments through the CM-managed dev/workflow/generate_workflows.sh directly (no GitLab CI/CD pipeline, no rocotostat.py instrumentation), so EXPDIR layouts and log artifacts differ from anything the forked branch's CI ever produces.

For the EIB MCP/GraphRAG, this means the existing collections — ingested from a snapshot that predates v17 additions — silently miss v17-only entities and return stale answers when asked about paradigm-specific behaviour. This document inventories what changed so the knowledge base can be updated deliberately rather than reactively.

HEAD = fd237a4f9 (dev_gsiupd2)
Merge-base with forked develop = 94548bc4f
Ahead/Behind = 21/0
Remote = github.com/CatherineThomas-NOAA/global-workflow

The 21 commits on top of forked develop are mostly straight develop progress (NSF/NCAR Derecho support, Ursa CI HPC account update, GFS low-res test case, ufs-model + GDAS hash refresh, complex2 GRIB2 packing, postamble executable, broad ush hygiene cleanup) plus the GSI-specific bump that names this branch (#4780, #4729). All ten submodules of the local dev_gsiupd2 clone have been initialized for this assessment.

derecho.yaml selects pbspro scheduler, queues develop/main, paths under /lustre/desc1. Resolutions C48–C1152 are flagged supported.

dev/workflow/applications/sfs.py — SFSAppConfig(AppConfig):

• Reads RUN (defaults to sfs) and NMEM_ENS from config.base.
• _get_app_configs builds stage_ic → fcst → atmos_products plus, when ensemble size > 0, efcs and atmos_ensstat; conditional wave block (waveinit, wavepostsbs, wavepostpnt, with bnd / pnt / bull variants); oceanice_products, prep_emissions, extractvars, arch_tars/globus, arch_vrfy, cleanup.
• _update_base forces RUN='sfs'.
• get_task_names returns the ordered task list mirroring the configs.

Companion config tree: dev/parm/config/sfs/{config.base.j2,config.efcs, config.arch_tars.j2,config.atmos_products,config.atmos_ensstat,...}.

dev/workflow/rocoto/tasks.py adds the _EXCLUSIVE resource family:

• New keys: clusters_exclusive, partition_exclusive, queue_exclusive, constraint_exclusive (validated from base).
• When a task config sets is_exclusive, the generator uses the _exclusive values (falling back to batch) and emits PBS :exclhost / Slurm --exclusive natives. Local exclusive_task variable consolidates the two scheduler branches.

This is the single entry point for v17 testing — it bypasses GitLab CI entirely and shells out to setup_expt.py + setup_workflow.py per case. New flags relevant to the v17 workflow:

Case selection block at line 138 (case "${option}"); ambiguous-selection guard at line 323; YAML iteration at line 640.

JGEFS_WAVE_STAT*, JGLOBAL_AERO_ANALYSIS_FINALIZE, JGLOBAL_ANALYSIS_STATS, JGLOBAL_ARCHIVE_TARS, JGLOBAL_ATMOS_SFCANL, JGLOBAL_ENKF_SELECT_OBS, JGLOBAL_FORECAST, JGLOBAL_MARINE_ANALYSIS_FINALIZE and _LETKF, JGLOBAL_WAVE_POST_SBS lose now-unneeded prod_util / module reset lines so that jobs run identically under rocoto and ecflow.

config.arch_tars and config.fetch migrated to Jinja2 templates across gcafs, gfs, gefs, and sfs, consistent with the broader move to template-everywhere configuration.

The DA-paradigm content (the GSI updates that name this branch and the synchronized ufs-model / GDAS bumps) lives inside the submodules, not in the global-workflow superproject.

Nine experiments, all Done per gw_cistat:

Layout per case (note: configs are flat at EXPDIR root, no config/ subdir):

EXPDIR/<case>/
  PSLOT.xml          # generated rocoto workflow
  PSLOT.db           # rocoto sqlite state
  PSLOT_lock.db      # rocoto lock
  PSLOT.scron.sh     # cron driver
  PSLOT.crontab      # crontab fragment
  config.*           # ~80 config files at root
  logs/
    YYYYMMDDHH.log   # per-cycle rocoto log
    scron.log        # cron driver log

There are no *_rocotostat.log files in v17 EXPDIRs because the forked-repo CI's dev/ci/scripts/utils/rocotostat.py wrapper is not part of this development line.

Sampled task topology (C96C48_hybatmDA_v17_gsiupd3, 46 unique task names, 10 metatasks) confirms the canonical three-cycledef pattern from the v16 diagram: gdas_half (single bootstrap) → gdas (cycled, 6h) → gfs (forecast, 24h). Task families:

• GDAS deterministic — gdas_prep, gdas_anal, gdas_analcalc, gdas_analdiag, gdas_sfcanl_{gcycle,regrid}, gdas_atmanlprod/upp, gdas_fcst_seg#seg#, gdas_atmos_prod_#fhr_label#, gdas_fit2obs, gdas_verfozn, gdas_verfrad, gdas_vminmon, gdas_anlstat, gdas_arch_tar_#tartype#, gdas_arch_vrfy, gdas_cleanup.
• EnKF (enkfgdas_*) hybrid DA — eobs, ediag, eupd, ecen#grp#, epos#grp#, esfc_{gcycle,regrid}, echgres, fcst_mem#member#, stage_ic, earc_*, cleanup.
• GFS deterministic — gfs_prep, gfs_anal, gfs_analcalc, gfs_atmanlprod/upp, gfs_fcst_seg#seg#, gfs_atmos_prod_#fhr_label#, gfs_tracker, gfs_genesis, gfs_arch_*, gfs_cleanup.

Throttling per workflow XML: <workflow scheduler="slurm" cyclethrottle="3" taskthrottle="25">, partition batch, cluster c6, account gfs-cpu.

Current knowledge base state (per get_knowledge_base_status):

Gaps when interrogating v17 content today:

1. applications/sfs.py and SFSAppConfig — missing from graph; find_callers_callees on these symbols returns nothing.
2. Derecho host (hosts/derecho.yaml, env/DERECHO.env, gw_*.derecho.lua, versions/*.derecho.ver) — not represented; get_workflow_structure will not list Derecho among supported platforms.
3. tasks.py is_exclusive resource branch — graph carries the older signature; queries about exclusive scheduling fall through.
4. ush refactor (postamble executable, err_exit, strict helpers, parsers, new helpers) — stale; lookups by file path drift (we already saw workflow_xml.py vs rocoto_xml.py).
5. gsi_enkf.fd feature/llvm branch content — never indexed (submodule snapshots are pinned to whatever was previously ingested).
6. generate_workflows.sh -S/-C/-I flags — search returns the older parameter set.

Captured here as the seed for the SDD descriptor; details to be elaborated in the spec phase.

1. Dual-branch ingestion. Add a second snapshot of dev_gsiupd2 with all submodules at v17 pins. Suggested collection naming: code-with-context-v17-dev_gsiupd2-<sha>, global-workflow-docs-v17-dev_gsiupd2-<sha>. Keep the existing develop collections untouched.
2. Branch / commit metadata. Tag every Chroma chunk and every Neo4j node with branch, commit_sha, and paradigm (develop | v17_pre_coupled). Expose a filter argument on search_documentation and on graph queries (analyze_code_structure, find_callers_callees, trace_execution_path) to scope answers to a paradigm.
3. Schema extensions for v17-only entities listed in §5. In particular: add the is_exclusive resource axis to the Task node, register the SFS application class, model the Derecho host the same way other hosts are modeled, and recognize the .j2 template family for archive/fetch configs.
4. Path-drift reconciliation. Sweep the graph for files renamed or removed since v8 ingestion (e.g. ush/fv3gfs_remap_weights.sh, workflow_xml.py) and emit a deprecation/redirect map so older queries resolve to current paths.
5. Cross-cutting collections stay shared. community-summaries, ee2-standards-v5-0-0-enhanced, ci-test-cases-v1-0-0 are paradigm- agnostic and need no split. Only code/docs collections fork.
6. Reference deliverable. This document, committed under docs/source/v17_paradigm_assessment.md, is the inaugural v17-aware reference doc and the input to the SDD spec phase.

• Detailed Neo4j schema deltas (node labels, properties, edge types).
• Chroma collection naming convention and migration rules.
• Ingestion pipeline modifications (paradigm-aware embeddings, dedup with the develop snapshot).
• Validation criteria and regression suite (sample queries that must be answered correctly per paradigm).
• MCP tool API surface for branch filtering (parameter naming, default behavior, backward compatibility).

• Direct inspection of both repos via the workspace file tools.
• git diff --stat fork_develop -- <paths> between the two checkouts (forked develop was fetched into the v17 clone as fork_develop for the comparison).
• gw_cistat plus per-case XML grep for task / metatask / cycledef counts.
• The EIB MCP gateway was deliberately not used as the source of truth for v17 content because its v8 collections predate the additions catalogued here. The gateway was used only for general Rocoto background and to confirm the v8 graph's path drift.