R8 Legacy Retirement — Design¶

Status: picked up by owner instruction (2026-09-29); scope authorized, slices S0–S5 ordered. Date: 2026-09-28 Branch: chore/r8-legacy-retirement (off main at 51f635e5) Supersedes: nothing. This is 12-postponed-roadmap.md R8 moving from the deferred index into its own spec, which is that file's own convention at :10.

Evidence tags follow the repo convention (AGENTS.md): [V] verified this session by source read at the named tip; [C] computed; [O] owner statement; [A] agent-reported, not independently verified. Keep every marker when editing.

0. Why this is open now¶

R8's trigger in the index is "Go-live holds 14 days". That trigger is reversed by owner instruction [O] — the owner picked it up on 2026-09-29 with two rulings that set the shape of the work:

• "we do not have to support backwards compatability" and "we should hard raise" [O]. PIPELINE_MODE=legacy must RAISE at boot. No deprecation window, no alias, no compatibility shim, no one-release warning. The precedent is ledger item 20 on the re-ID swap, whose wording is nearly identical.
• No A5500. Repo + GB300 only. S1's 9,722 MiB and S4's 16.2 s belong to that box (ledger items 40–41) and are neither re-measured nor restated as repo claims here.

The reason the deletion matters functionally rather than cosmetically: the legacy path is reachable by environment variable today. backend/core/config.py:1084 accepts "legacy" and :1086 only logs a warning [V]. A deployment with a stale .env boots the unsupported pipeline.

The owner's stated motive is agent-experience, not code hygiene: "so we dont confuse future agents."

1. What goes¶

Everything below verified [V] at main 51f635e5, which is the tip this spec was written against. The measured detail — every line number, importer and trap — lives in docs/plans/2026-09-28-r8-legacy-retirement-scope.md.

2. What stays — and the one extraction that must happen first¶

nemotron_analyzer.py is ~98% legacy, but :294-411 is imported by the shipped VLM path [V]: vlm_client.py:65-69, vlm_analyzer.py:74, evaluation/vlm_replay.py:56, and main.py:773 inside the live constrained-decoding startup gate. Symbols: ConstrainedDecodingNotEnforced, _is_length_truncated, build_probe_schema (and VerificationRowOutcome, PROBE_PROMPT, _TRUNCATED_STOPS, _probe_completion). They move to backend/services/constrained_decoding.py before the file is deleted; deleting whole breaks vlm at import.

The three VLM-reachable loaders stay, with model_loader_base.py: face_recognizer_loader, osnet_loader, fast_alpr_loader [V]. The other 19 are reachable only through enrichment_pipeline, model_zoo._LOADER_MAP, or prompts.py under if TYPE_CHECKING: — which erases at runtime and is therefore not evidence of life.

This is not a judgement call about capability. core/container.py:508 builds the enrichment collaborators only when pipeline_mode == "legacy", and no API route constructs EnrichmentPipeline [V]. Retiring legacy makes the tier unreachable by construction.

Stays for a different reason than "it works": the specialists are lookups against a store the pixels cannot reach (face→known_persons, plate→registered vehicles, re-ID→PersonEmbedding via household_matcher.load_person_gallery). The VLM receives the raw key frames as base64 (vlm_client.py:499) and perceives the scene itself, which is why Florence and the attribute zoo are substitutions rather than additions. If more signal is wanted later, the category worth adding is detection/lookup — the sanctioned optional threat specialist (GATEWAY_ENABLE_THREAT, residency.py:69) — not a second perception model.

Other must-not-break [V]:

• notification_filter.py:35 should_notify — implemented, unwired. M1's exit; a separate job.
• threat_detections (enrichment.py:100) — read by alert_engine.py, api/schemas/alerts.py.
• household.py:155 PersonEmbedding — LIVE, while enrichment.py:218 reid_embeddings is legacy.
• The llama-cache named volume (prod.yml:1621) — mounted by ai-vlm at :262.
• GPU_LLM, and the CTX_SIZE/PARALLEL pair — shared with vlm's slot budget (32768/2).
• The models.yml row carrying triton_name: reid — one row feeds the CPU OSNet loader and the Triton repository. Keep the whole row.
• Triton yolo26, reid, and the opt-in threat — S1's PASS was measured against exactly yolo26+reid (residency.py:65, threat off by default).

3. Slices¶

One slice, one PR, TDD red-first. Order matters where noted.

Ordering argument for S4 early [V reasoning]: select_preload_candidates (main.py:644) preloads by enabled AND preload and is not keyed to mode, so smoke-fire-yolov8n's 350 MB preloads in shipped vlm mode today. S1's PASS was measured with that dead weight inside the budget, so retiring it moves S1 down — headroom, never risk. Table retirements are also smaller and more reversible than deleting a module the DI graph still names.

4. Triton residency¶

FULL_MODEL_SET (residency.py:46-60) holds 14; VLM_MODEL_BASE is ("yolo26","reid") (:65) with threat opt-in (:69). The 11 others are legacy-only, so pruning them cannot move S1's number. Face (SCRFD + w600k_r50) is CPU-side in the backend and is not in Triton at all. Tests that pin the pair move with the change: test_gateway_model_set_compose.py, ai/gateway/tests/test_residency.py, test_entrypoint_residency.py.

5. Schema¶

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/. So a table retires as: delete the model class, add a dated DROP file. Precedents: 2026-09-26-face-vector-provenance-model-id.sql, NEM-5051-convert-alert-json-to-jsonb.sql.

6. Frontend¶

R8 :108 names "their frontend panels" as in-scope, and they are real [V]: components/enrichment/EnrichmentViewer.tsx, components/events/EnrichmentBadges.tsx, EnrichmentPanel.tsx, EventEnrichmentSummary.tsx, utils/poseVisualization.ts. The tree is frontend/ — not src/frontend/, a path several scoping agents asserted and none exists.

7. Minefield¶

• Clear stale __pycache__ before trusting a deletion guard (R8 :118: a stale backend/api/helpers/__pycache__ false-reddened one on 2026-09-23).
• Coverage gates are the merged floors — 84 unit (ci.yml:613), 37 integration (ci.yml:1215), not pyproject.toml's 85. Report the direction; never lower a floor.
• Deleting test files moves the suppression census (ci.yml:101-104 vs .github/suppression-baseline.json). Regenerate via scripts/suppression-registry-gen.py; never hand-edit a generated file.
• ~96 legacy-only test files go with their module; ~61 more merely mention legacy and are edited, not deleted.
• 66 of those 96 error at import in this sandbox (no libxcb.so.1/cv2). Environmental — judge by CI.
• Never rewrite frozen research prose, and never edit a spec's Status/rev line: owner-only.
• docker-compose.ghcr.yml has no ai-vlm service at all. Report it; do not fix it inside a retirement slice.

8. Open, owner's call¶

1. models.yml: delete legacy rows, or keep enabled: false as provenance?
2. Does R8 reach the ai/enrichment / ai/enrichment-light providers, or backend only?
3. Keep full as a Triton dev affordance? (Default: no — hard-raise GATEWAY_MODEL_SET.)
4. nemotron_url has live, non-mode-gated callers [V]: summary_generator.py:84 (its scheduler at main.py:1092 is gated on Redis only, not mode) and prompt_service.py:938 (POSTs {_llm_url}/completion; router registered at main.py:1649). Both already degrade to warnings because ai-llm is profile-legacy — so deleting makes an existing breakage permanent. Re-home them on the VLM, or retire the features explicitly.
5. Whether keeping the threat detector switched-on is wanted now (costs S1 VRAM — re-measure).