Skip to content

Commit 74bff46

Browse files
fix(plan): reconcile seven retrieval-first model incompatibilities
Keep the CURRENT Nemotron embedding provider, retire the two general Ollama rows and their hidden mise/functions/dispatch paths, and replace QMD's mandatory 0.6B vector prerequisite with the existing CPU lexical role and document gate. Preserve historical assets and observations. Missing current state needs_owner; no runtime, model, index, unit or host is bootstrapped by this repair. This is source-only for grand-catalog #810, after frozen #723. Check-plan against #810 db93014 passes; the authentic combined source passes 129 touched tests with 3 existing skips, including unchanged-copy. Root-source validation passes after temporary native hash registration; downstream rebuilds its registry last. SOTA: db930144:docs/decisions/2026-10-06-retrieval-first-local-models.md:27-33,58-61,115-118 SOTA: https://huggingface.co/nvidia/Nemotron-3-Embed-8B-BF16/blob/d1f2f25730bbd775b99b29185134bc86653bf2d1/README.md SOTA: vllm-project/vllm@db9527a:docs/serving/online_serving/openai_compatible_server.md SOTA: tobi/[email protected]:README.md (native named-index lexical search) SOTA: 979291de:evidence/artifacts/new-wsl-install-plan-20261002/check_plan.py:31-77,431-450,665-678
1 parent 979291d commit 74bff46

9 files changed

Lines changed: 363 additions & 657 deletions

File tree

‎evidence/artifacts/new-wsl-install-plan-20261002/README.md‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,8 @@
11
# New WSL install plan
22

3-
Current structural inventory (2026-10-05): 84 rows, 192 install commands and 113 acceptance entries (including two additional checks). The source includes 66 configuration assets. D-ollama-2 carries the updated owner-approved user unit and removes the warmup asset; #713 supplies the additional rows. These counts describe the source, not host acceptance.
3+
Current structural inventory (2026-10-07): 84 rows, 183 install commands and 108 acceptance entries (including two additional checks), with 61 installed rows, two measurement-only rows and 21 exclusions. The source includes 66 configuration assets. The retired Ollama unit and environment example remain historical assets and are never installed. These counts describe source integration with #810's settlement data, not host acceptance.
4+
5+
The retrieval-first ruling supersedes the older local-model recipes below. The selected embedding row checks the existing Nemotron 8B provider on loopback 28231, including its exact returned model name and two finite, normalized 4096-dimensional vectors. It bootstraps no runtime or weights; an absent provider is nonzero `needs_owner`. General local generation and its Ollama server have empty executable commands and acceptance, and Ollama is absent from mise. QMD's selected CPU lexical index replaces the compulsory 0.6B/vector prerequisite and must return the known harness document. The prior recipes, owner-unit history and actual recorded runs remain unchanged as provenance. [Sources and boundaries](SOURCES.md#retrieval-first-plan-parity-2026-10-07) distinguish these integration checks from upstream tests and live acceptance.
46

57
**Run the plan from a checkout of the repository** (step F7, "Clone origin/main", of `adoption/platforms/linux-wsl2-new-distro.md`). `worktrunk` and the convergence validators change into `repo_root`, the checkout three levels above this folder, so `install.sh` and `accept.sh` stop with a clear message when `repo_root` is not a git checkout (`install.sh --list` needs none).
68

‎evidence/artifacts/new-wsl-install-plan-20261002/SOURCES.md‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1216,3 +1216,11 @@ This #723 source integration reuses the prior B2 metadata and preserves its fail
12161216
- The handbook's maintained current-output contract is native-stack@4e07a17d55860d3ae15c7437a348d59ff2fc40a1:docs/decisions/2026-10-01-new-wsl-handbook-generator.md:665 and tests/test_new_wsl_handbook.py:1672-1689. The generator-owned receipt retains the previous pair while binding its current outputs exactly; all prior validation observations remain unchanged.
12171217

12181218
J723c extends the existing test contracts, using CPython3.13's [Path.rglob](https://docs.python.org/3.13/library/pathlib.html#pathlib.Path.rglob) and [Path.is_file](https://docs.python.org/3.13/library/pathlib.html#pathlib.Path.is_file), matching this repository's check_plan.py recursive config reader at native-stack@4e07a17d55860d3ae15c7437a348d59ff2fc40a1:649. Nested drop-in contents remain part of port checks, and the exact root script set explicitly includes the frozen Inspector probe. The AgentsView current repin classification follows tests/test_agentsview_qualification.py:50-76,292-347; the older launcher fixtures remain dated predecessor inputs, not current pins. Native Git/Python/env commands no longer require RTK in CI or at runtime; optional parsing of earlier RTK-shaped client events remains supported. The already-landed currency update supplies the native git-grep registry recipe. Historical RTK observations and component-specific tests are retained.
1219+
1220+
## Retrieval-first plan parity (2026-10-07)
1221+
1222+
The seven plan/manifest incompatibilities are recorded in grand-catalog's unchanged-copy check of 2026-10-06: one embedding identity mismatch and three identity/install-state mismatches for each retired Ollama row. The selection source is native-agent-stack@db930144:docs/decisions/2026-10-06-retrieval-first-local-models.md:27-33,58-61,115-118, implementing the command center's final local-model ruling. This changes the executable plan and its necessary mirrors; it does not change the settlement, historical verdicts or observed results.
1223+
1224+
The kept embedding provider is the existing Nemotron 8B endpoint. [NVIDIA's model card at d1f2f257](https://huggingface.co/nvidia/Nemotron-3-Embed-8B-BF16/blob/d1f2f25730bbd775b99b29185134bc86653bf2d1/README.md) specifies full 4096-dimensional, L2-normalized embeddings and the exact `query: ` and `passage: ` prefixes, and demonstrates serving the named model through vLLM. [vLLM v0.31.0's native HTTP API](https://github.com/vllm-project/vllm/blob/db9527a46873454610df6dbedf79a36d6bf1a7f6/docs/serving/online_serving/openai_compatible_server.md) supports `/v1/embeddings`. The model card's installation example uses vLLM 0.25.0; it is not evidence that 0.31.0 was installed or accepted. The CURRENT 0.31 provider and loopback 28231 are owner-kept state from the ruling, checked through their native HTTP surface. Missing or mismatched state is nonzero needs_owner; this row adds no server, environment, model download, rebuild or fresh-host bootstrap.
1225+
1226+
Local generation and its Ollama server are explicitly excluded. The prior owner-approved unit remains byte-identical historical data (native-agent-stack@979291de:evidence/artifacts/new-wsl-install-plan-20261002/config/ollama.service:1, SHA256 5e8e8bc4…), alongside its existing environment example; neither is installed. The checker rejects installation prerequisites, operational commands, services, config placement and implicit mise Ollama installation for the retired rows. QMD 2.8.3 keeps its native lexical role under the selected named index with `QMD_FORCE_CPU=1` and inherited `INDEX_PATH` removed; its compulsory 0.6B GGUF/vector prerequisite is retired, without downloading or selecting a replacement embedding model. Native package/SDK execution is distinct from our synthetic API/retirement controls. These source changes and fixtures qualify plan integration only; the command center owns the host apply and smoke.

0 commit comments

Comments
 (0)