R8 legacy retirement — measured scope (2026-09-28)¶

Scope notes for the retirement slices, written on the GB300 dev partition [V: this sandbox is the agent partition per ledger E10]. Nothing here has been executed — no file has been deleted, no test tier run. Every line reference was read at the tip of feat/vss-gaming-gpu-profile (1f2921a4) and re-checked by hand; the parts a 7-agent scoping pass asserted and I could not reproduce are marked as such. It is the detail layer for the goal prompt at docs/plans/2026-09-29-r8-legacy-retirement-goal-prompt.md; that prompt is the order of work, this file is the evidence.

Owner rulings that set the scope (both 2026-09-29, recorded here before the work):

• "we do not have to support backwards compatability" and "we should hard raise" — so PIPELINE_MODE=legacy must RAISE at boot, not warn. This matches ledger item 20's near-identical wording on the re-ID swap ("full swap. ignore previous architecture…").
• The R8 trigger is reversed by instruction: 12-postponed-roadmap.md:23 defers this deletion to "Go-live holds 14 days"; the owner is picking it up now, per that file's own convention at :10 (a picked-up item gets its own dated spec and a mark here).
• No A5500. Everything below is repo + GB300 work. The pinned PASS numbers (S1 9,722 MiB of 20,890; S4 p95 16.2 s of 30 s, ledger items 40–41) belong to that box and must not be re-measured or re-stated as repo claims.

1. What is actually reachable today [V]¶

The legacy path is live-by-environment-variable, so its removal is a functional change, not a cosmetic one.

• backend/core/config.py:1061-1090 — the validator's own comment says "legacy" parses only because its code stays in the repo until R8 deletes it. :1084 accepts the two spellings, :1086 logs a warning. It does not reject.
• Consequence: a deployment with a stale PIPELINE_MODE=legacy in its .env boots today and runs the unsupported pipeline. Under the hard-raise ruling it must crash at boot.

1a. Two shared config pairs that are NOT legacy-only [V]¶

• GPU_LLM is the LLM's and the VLM's GPU: docker-compose.prod.yml passes it to ai-llm (:145,:159,:209) and to ai-vlm (:251,:266,:35x). Deleting the ai-llm block must not delete the variable or its .env.example row.
• CTX_SIZE / PARALLEL is one env pair feeding two different slot budgets: config.py binds nemotron_context_window→CTX_SIZE (legacy, 262144) and llama_slot_count→PARALLEL, while the shipped vlm slot uses 32768/2. Removing the legacy defaults must leave the vlm binding untouched — this pair is exactly where a careless deletion changes the measured slot ceiling behind S1's back.
• nemotron_url has live, non-mode-gated consumers beyond the legacy pipeline: summary_generator.py:84, prompt_service.py:679, pipeline_quality_audit_service.py:136, performance_collector.py:475. Whether those features are live under vlm mode is not established here — §11 item 5.

2. The legacy branch sites [V]¶

Instrument (wide grep — the narrow one is a trap, see below):

grep -rn '"legacy"' backend/ --include=*.py | grep -v tests   # 13 lines at the tip

Trap: the narrower grep -rn 'pipeline_mode == "legacy"\|pipeline_mode != "legacy"' returns 7 and silently misses backend/services/pipeline_factory.py:52, which branches on a local mode variable assigned at :51 (mode = _settings_for_mode().pipeline_mode). An agent that enumerates with the narrow grep ships a live legacy branch.

Sites, all [V] at the tip: main.py:733, main.py:1134; pipeline_factory.py:52 (the module is 59 lines and is essentially the mode switch itself); system_broadcaster.py:1028; container_orchestrator.py:536; core/container.py:508; api/routes/health_ai_services.py:111; api/routes/system.py:1164 (+ the legacy health body it short-circuits); api/schemas/enrichment_data.py:442,454; config.py:1084,1086.

pipeline_workers.py:800 is only a comment — its real coupling is the from backend.services.nemotron_analyzer import NemotronAnalyzer at :67.

3. nemotron_analyzer.py is NOT wholly legacy [V — this was the pass's most important catch]¶

backend/services/nemotron_analyzer.py is 5,429 lines; ~98% is legacy. But :294-411 is imported by the VLM path, verified at each importer:

main.py:773 is inside the live constrained-decoding startup gate — the comment at main.py:769 says the VLM mode's gate "shares the same shared ConstrainedDecodingNotEnforced verdict vocabulary". Deleting the file whole therefore breaks the shipped pipeline at import, and the S-1 enforcement verdict with it.

Slice order: 2a hoist :294-411 (also carries VerificationRowOutcome:294, PROBE_PROMPT:323, _TRUNCATED_STOPS:335, _is_length_truncated:338, _probe_completion:352, build_probe_schema:396) into backend/services/constrained_decoding.py, repoint the four importers, red-first with an import test. 2b deletes the ~5,321 remaining lines.

Adjacent legacy-only modules — sizes now wc-verified [V]: enrichment_pipeline.py 7,662 (the pass's "3,334" was enrichment_client.py's number, swapped), enrichment_client.py 3,334, nemotron_streaming.py 575, analyzer_facade.py 266, prompt_auto_tuner.py 205, nemotron_latency_optimizer.py 650, plus two the first pass missed: skeleton_action_service.py (only importer: enrichment_pipeline) and pose_analysis_service.py (zero non-test importers — already orphaned). DI/type bindings naming NemotronAnalyzer also live at ai_fallback.py:47,238 (+health use :433), api/routes/events.py:100,105,2584, pipeline_workers.py:67, api/dependencies.py:346,949,1430, core/dependencies.py:60,100 — several are TYPE_CHECKING-only imports (from backend.services... import under if TYPE_CHECKING:), which erase at runtime: they must be edited, but deleting the module does not crash them at import — distinguish those from the eager ones (events.py:100 is a real runtime import).

Safe to delete without touching M1: nemotron_analyzer.py has zero hits for should_notify|NotificationFilter|notification_filter [V], so the unwired M1 rule at notification_filter.py:35 survives the deletion untouched.

4. Enrichment and the specialists do not overlap [V]¶

backend/services/vlm_specialists.py has zero references to enrichment — the face, plate and re-ID legs are built from the module's own handles (collect_face_text:499, collect_plate_text:552, collect_reid_text:645). Deleting enrichment_client.py (3,334 L) and enrichment_pipeline.py cannot orphan the specialists.

But the package init is the sever-most cut [V]. backend/services/__init__.py eagerly re-exports both targets with no try/except: EnrichmentPipeline et al. at :95 and NemotronAnalyzer at :245. Deleting either module without first editing that file breaks import backend.services everywhere at once — the whole tier, not one caller. Edit the init in the same slice, ahead of the rm.

Other shared infra that is not legacy and must survive [V]:

• llama-cache named volume (docker-compose.prod.yml:1621) is mounted by ai-vlm at :262 (and ai-llm at :152). Deleting the ai-llm service block must not delete the volume — that takes the VLM's CUDA/JIT caches with it.
• The three specialist loaders stay: face_recognizer_loader.py, osnet_loader.py, and the plate leg. osnet_loader reads its weights and pin from the models.yml row that also carries triton_name: reid — one row feeds the CPU in-process loader and the Triton repository. Keep the whole row; the row is entangled, not just the Triton half.
• Florence is legacy-reachable, not VLM-reachable [V] (this corrects an earlier, too-strong draft line that put prompt_service.py on the live list — its only Florence token is the docstring example at :712). The real Florence runtime importers are services/__init__.py:136,145 (the eager re-export — the sever-most cut, above) and model_zoo.py:66 (the _LOADER_MAP wiring). florence-2-large is already enabled: false; florence-2-base is the Triton row. Trace the wiring, don't trust R8's roadmap line — see the loader census below.
• ai/gateway/main.py:60,66 builds ALL_MODELS from list(FULL_MODEL_SET); pruning FULL_MODEL_SET must move that and the /health reporting together, or the gateway advertises models its repository no longer serves.
• api/routes/model_management.py:162 ROUTER_SUFFIXES = ("/enrichment", "/enrich-lt") — the live model-readiness API is built on the two enrichment routers, so the frontend readiness surface breaks the moment those routers go. Re-scope the endpoint, don't just delete.
• models.yml is one file, many readers (setup_lib/model_downloader, services/model_zoo, gateway residency, the precheck) on three different filters. A row's removal has four blast radii; treat the file as the shared surface it is.

4a. The keeps/goes census: the enrichment tier is legacy-only by construction [V]¶

Answering the owner's "we keep the VLM and the supplemental models, right?" — yes, but the reason is stronger than a list of keepers, and it is what makes R8's :109 phrase "the in-process loaders the VLM path no longer calls" true rather than hopeful.

The construction gate. core/container.py:508 builds the analyzer's enrichment collaborators only inside if get_settings().pipeline_mode == "legacy": — context_enricher and enrichment_pipeline are passed as kwargs then and never otherwise (vlm_analyzer does not take them). So once legacy is gone, EnrichmentPipeline has no construction site and no API route reaches it (grep enrichment_pipeline backend/api/routes/ = 0 hits [V]). The whole enrichment tier is not "decided to be legacy" — it becomes unreachable the moment S1 lands. R8's "What stays: the face, re-ID and plate specialists" (:114) is therefore an accurate, not merely aspirational, line.

The loader census — all 22 backend/services/*_loader.py, classified by real (non- docstring) importers, marking VLM-path reachability [V]:

Two traps inside that census:

1. prompts.py is not evidence of life. Its loader imports at :674-691 sit under if TYPE_CHECKING: — type-only, erased at runtime. An import grep that doesn't exclude the TYPE_CHECKING block reads 15 legacy loaders as live consumers. prompts.py's real importers (nemotron_analyzer, nemotron_streaming, analyzer_facade, enrichment_pipeline) are all legacy-side.
2. threat_detection_loader is legacy but the Triton threat model is not. Two different implementations, same word. Keep residency.py:69 VLM_THREAT_MODEL and the models.yml:309 triton_name: threat row; threat_detection_loader goes with the tier.

Preload residency, so the S1 number is understood [V]: BACKEND_MODEL_PRELOAD (config.py:2886, default false) makes main.py:644 select_preload_candidates load every enabled row that declares preload: true — exactly four: osnet-ain-x1-0 (100 MB, F11), face-detector-scrfd (0), face-recognizer (0), and smoke-fire-yolov8n (350 MB, enrichment-only — imported only by enrichment_pipeline/model_zoo/nemotron_analyzer). So the shipped VLM path preloads three GPU legs, and the fourth preloaded model is dead in vlm mode. Its row's comment even says "CRITICAL: never evict, preload at startup", which select_preload_candidates' own docstring records as "decorative" — never_evict has no consumer. Retiring smoke-fire means removing that row's preload: true too, or a future agent re-reads it as resident.

5. Triton residency — there is no S1 risk here [V]¶

This section exists because an earlier draft of this file asked the owner an ill-posed question about "moving a Triton row". The facts dissolve it:

• ai/gateway/residency.py:65 — VLM_MODEL_BASE = ("yolo26", "reid"); :69 adds threat as an opt-in. Shipped state is GATEWAY_ENABLE_THREAT=false (docker-compose.prod.yml:482), so the shipped set is exactly yolo26 + reid, and that is the footprint S1's 9,722 MiB was measured against.
• FULL_MODEL_SET at :46-60 holds 14 models; 11 are legacy-only — clip, clip_text, florence2, vehicle, fashion_clip, demographics_age, demographics_gender, pet, depth, pose, stgcn_action. Pruning those cannot move S1's number.
• threat is not legacy — it is the vlm set's optional fourth. Keep it.
• Face is not in Triton at all. SCRFD + w600k_r50 run CPU-side in the backend, loaded by face_recognizer_loader (vlm_specialists.py:328). Do not look for them in the gateway.
• Tests that pin the pair and must move with the change: backend/tests/unit/core/test_gateway_model_set_compose.py, ai/gateway/tests/test_residency.py, ai/gateway/tests/test_entrypoint_residency.py.

Open, low-stakes: once legacy is gone, get_model_set("full") serves only a deliberate "bring the whole repo up" dev run (residency.py:26). Default is to delete it and make GATEWAY_MODEL_SET hard-raise on anything but vlm, consistent with the no-backcompat ruling; keeping it as a dev affordance is the owner's call.

6. Tables — no Alembic, and two tables that look alike but are not [V]¶

There is no Alembic in this repo: schema comes from backend/scripts/init_schema.py + core/database.py, with hand-authored dated SQL in docs/api/migrations/ (precedents: 2026-09-26-face-vector-provenance-model-id.sql, NEM-5051-convert-alert-json-to-jsonb.sql). So retiring a table = delete the model class + add a dated DROP SQL file.

• backend/models/enrichment.py is not a wholesale delete. pose_results:47, demographics_results:160, action_results:265, reid_embeddings:218 are legacy-only — reid_embeddings is written only by enrichment_pipeline.py:5848.
• threat_detections:100 stays — read by services/alert_engine.py and api/schemas/alerts.py.
• The name trap: enrichment.reid_embeddings is legacy, but backend/models/household.py:155 PersonEmbedding is LIVE — the VLM re-ID leg loads its gallery through household_matcher.load_person_gallery:211, which is deliberately the single reader so the one-embedding-space rule (F11/item 20) holds exactly once. Also live: face_identity.py:80 known_persons, :139 face_embeddings.

7. ai_contract is the real gate on deleting ai/ [V]¶

The retired serving dirs (ai/florence, ai/enrichment, ai/enrichment-light, ai/clip, ai/triton) are referenced from backend/ai_contract/ (operations.py, providers.py, fake/generators.py) and backend/tests/contracts/ai_providers/*. That contract tier is PR-gated (Contract Tests (API Schema) runs uv run pytest backend/tests/contracts/ -n0 --timeout=30, .github/workflows/ci.yml:1363).

Named first: operations.py:380-391 declares an llm_completion Operation whose evidence string cites nemotron_analyzer.py:465, with schemas/llm_completion.{request,response}. That row must retire in the same slice that deletes the analyzer, or the tier goes red. Conclusion: retire one provider per slice, never rm -rf ai/<dir>.

8. Test and CI blast radius [A where not marked V]¶

• Coverage gates are the merged floors, not pyproject. ci.yml:613 fails under 84 (unit), ci.yml:1215 under 37 (integration) [V]. pyproject.toml:636 fail_under = 85 is the local gate; the unit shards run --cov-fail-under=0 and only collect (ci.yml:430-467) [V]. Deleting legacy code moves the denominator in a direction nobody has measured — measure per slice, report the sign, never lower a floor.
• Legacy-only test mass: ~96 files / 86,970 lines / 3,175 AST test defs (89 unit files / 3,050 defs). ~61 further files merely mention legacy and must be edited, not deleted (example given by the pass: test_model_warmup.py, 41 nemotron refs, subject is warmup).
• The suppression census, not a "test-count mirror." The 99→101 in PR 6698's own commit messages is category pytest_skip_imperative of the escape-hatch census, adjudicated by ea79a858 and gated at ci.yml:101-104 against .github/suppression-baseline.json, regenerated by scripts/suppression-registry-gen.py [V]. Deleting test files changes that count → regenerate, never hand-edit. (An earlier draft called this a test-count mirror; that was wrong and the pass corrected it.)
• Environmental noise: 66 of those 96 files error at import in this sandbox for want of libxcb.so.1 (cv2 is absent — reproduced). Judge deletions by CI, not local runs.
• No test anywhere asserts a deleted module is absent [A — the pass grepped for it and found only the stale-pycache scan at scripts/a5500_precheck.py:36]. So the grep-gate in S1 is a test the implementer must author.

9. Docs a future agent reads [V for path existence]¶

• Currency, ship alone: dated in-doc banners over docs/vss-integration/README.md:6 and AGENTS.md:6, both claiming "nothing implemented yet" — false since 4bfd6fa4 made vlm the shipped default (docker-compose.prod.yml:578). Precedent to imitate: 06-repo-a-readiness.md:11 "Retirement note — 2026-09-21 [V]". Also 00-context.md:5 ("Research in progress"), README.md:18-19 (credits the Brev matrix with measurements that never ran, ledger :148), and ledger :3's rev-5 cite (spec is rev 7). Never rewrite frozen research prose, and never edit a spec's Status/rev line — that is owner-only house convention, not drift.
• Post-deletion: root README.md/AGENTS.md, ai/AGENTS.md, backend/AGENTS.md, mkdocs.yml nav, docs/developer/{moe-offloading,llm-inference-optimization,multi-gpu}.md, docs/reference/{models,nvidia-technology-inventory}.md, llms.txt. (Two paths an earlier draft asserted do not exist: docs/developer/nvidia-technology-inventory.md is actually docs/reference/…, and there is no ai/llm/ — the build context is ./ai/nemotron at prod.yml:139.)
• docker-compose.ghcr.yml has no ai-vlm service at all — it cannot serve the shipped mode today. Report it; do not "fix" it inside a retirement slice.
• docs/superpowers/plans/2026-09-27-a5500-vlm-bringup-checklist.md:42-44 is GENERATED and already stale. If a slice breaks a generator assertion, report it — never hand-edit a render (scripts/a5500_precheck.py is the authority).

10. Known-truth hazards carried forward¶

• Ledger items 39–41 cite commit hashes (2fd5c54bc, d7b3fd3c6, cedac0a64, …) that do not resolve as git objects at this tip — a branch rewrite, not fabrication. A hash grep reads the newest evidence as invented; it isn't.
• The newest S4 number is 16.2 s (item 41), superseding item 39's 21.7 s.
• origin in this sandbox is the dead host path /home/msvoboda/github/…; the real branch is on remote github, and gh needs --repo mikesvoboda/nemotron-v3-home-security-intelligence.
• Do not delete notification_filter.py:35 should_notify. All three verdict rules are implemented; only the wiring is missing. That is M1's exit, a separate job.

11. Not decided here (owner's call, with the recommended default)¶

1. models.yml: delete legacy LLM rows, or keep enabled: false as provenance? The X-CLIP precedent kept them and calls the file "owner-owned, untouched". Default: ask before deleting.
2. Does R8 reach the enrichment AI providers (ai/enrichment, ai/enrichment-light) or only the backend client + pipeline + legacy LLM? R8's own text at :108-115 reads wider than the ledger's description. Default: backend first, providers as their own slice.
3. Keep full as a Triton dev affordance? Default: no — hard-raise GATEWAY_MODEL_SET.
4. Whether R8's "(including the 0.3 changes)" imposes anything beyond deleting the file.
5. nemotron_url is read by summary_generator.py:84, prompt_service.py:679, pipeline_quality_audit_service.py:136 and performance_collector.py:475 — none of them mode-gated. Two of the four are confirmed live callers, not just readers [V]: summary_generator.py:84 (the hourly dashboard summary — its scheduler start at main.py:1092 is gated only on redis_client is not None, NOT on mode) and prompt_service which POSTs {_llm_url}/completion at :938 (registered router: main.py:1649). Because ai-llm is profiles: [legacy] (prod.yml:125), it is already down in the shipped vlm mode, so both features degrade to warnings today (summary_job.py:252-255 catches TimeoutError) — deleting the LLM makes a currently-incident breakage permanent. Owner: is a vlm-backed summarizer wanted (point it at ai_vlm_url), or is the feature retired with the LLM? Decide before S2; do not silently orphan it.
6. ~~Florence's live non-mode-gated consumers~~ — RESOLVED [V], see §4's census. Florence is legacy-reachable, not VLM-reachable; the only runtime importers are the eager services/__init__.py:136,145 re-export (a sever point, not evidence of life) and model_zoo.py:66. prompt_service's only Florence token is a docstring example (:712). Default: delete with the tier, after editing the __init__ re-export.
7. Does retiring the attribute zoo cost understanding the owner wants back? (owner asked 2026-09-29; recommendation here, decision is the owner's.) The VLM receives the raw key frames as base64 (vlm_client.py:499), so Florence-2 and the 19 attribute loaders re-perceive what the 8B already sees — re-adding them (even "with a more modern version") is a weaker parallel perception, not new information. The three survivors stay for a different reason: they are lookups against a store the pixels can't reach (face→knownpersons, plate→registered vehicles, re-ID→PersonEmbedding gallery via load_person_gallery). If the owner wants more signal, the category worth adding is _more detection/lookup, not more perception: (a) switch on the already-sanctioned threat specialist — GATEWAY_ENABLE_THREAT (prod.yml:482, default false), resident at residency.py:69, D3 rev-6's optional fourth; it is a precision/recall detector (not a describer) and the one legitimately worth modernizing off threat-detection-yolov8n (models.yml:309). Costs S1 VRAM — re-measure if enabled. (b) R11 glass-break audio is the other off-pixels signal, deferred (no consumer card fits the audio stack). Counter-evidence against re-bolting the zoo: S3 already fails 90% in every arm for scoring reasons, and moves ~3/20 between identical runs (ledger :403) — attribute text adds reconciliation surface below the noise floor, inside a 1024-token verdict budget (vlm_client.py:91). Default: proceed with the R8 deletion; treat the threat detector as a separate opt-in decision.

12. Reconciliation of the 14-agent scoping workflow (wf_5ccaa8ff-0c7) [V]¶

Every line below was re-read at the tip after the workflow returned. The workflow's claims were largely sound; a meaningful number of its paths and line numbers were not, so nothing was folded in without a re-read.

Refuter claims I REJECTED — the refuter was wrong, my original [V] stands:

• model_management.py:162 ROUTER_SUFFIXES = ("/enrichment", "/enrich-lt") exists (re-read verbatim; the two lanes are explained at :106/:121). §4 keeps it.
• main.py:733 if settings.pipeline_mode == "legacy": exists (re-read verbatim).
• ai_fallback.py is line drift, not fabrication: the NemotronAnalyzer binding is :47 and :238; the health-check use is :433 (the spine cites :434). Off by one, real symbol.

Where BOTH sides had the wrong path (the signal is real; the citation was not):

• The frontend is frontend/, not src/frontend/. Every workflow cite of the form src/frontend/src/components/enrichment-panel.ts:19-31 is a hallucinated path; my own first grep of src/frontend came back empty for the same reason. On the real tree the enrichment surface is 106 files, and the child-object renderers R8's :108 ("their frontend panels") actually names are real: components/enrichment/EnrichmentViewer.tsx, components/events/EnrichmentBadges.tsx, EnrichmentPanel.tsx, EventEnrichmentSummary.tsx, plus utils/poseVisualization.ts (draws COCO-17 skeletons from the retired pose keypoints).
• Four enrichment config fields exist and are dead in vlm mode, but at the real lines florence_url:1582, clip_url:1586, enrichment_light_url:1594 (+ enrichment_url, and the :1831 validator list) — not the :751-818 range the pass reported.

NEW, confirmed, must be in scope (missed by my first pass):

• backend/services/skeleton_action_service.py — imported only by enrichment_pipeline.py, zero VLM-path importers → legacy-side, retires with the tier.
• backend/services/pose_analysis_service.py — zero non-test importers at all (already orphaned) → retires with the tier.

Also corrected: enrichment_pipeline.py is 7,662 lines, not the 3,334 the scoping pass reported (that is enrichment_client.py's size — the two were swapped).

A sequencing argument worth adopting — do the table/row retirements FIRST [V reasoning]. select_preload_candidates (main.py:644) preloads by enabled AND preload from models.yml; nothing in it is keyed to pipeline_mode, so smoke-fire's 350 MB preload runs in shipped vlm mode today. S1's 9,722 MiB PASS was measured with that dead 350 MB inside the budget, so retiring smoke-fire moves S1 down — freeing headroom, never risking the bar (an owner-asked question this answers). The workflow's strongest single finding: it argues the table/row retirements belong ahead of the code deletions, not after — retiring a table's model class + dated DROP is a smaller, more reversible cut than deleting a module the DI graph still names. The spine keeps its one-slice-one-PR rule either way; this only reorders which slice goes first.