🥈 [SR-02] trainer-runner — E2E TTT O(1) per-chunk core

[gHashTag]

Soul-name: Chunk Champion · Codename: LEAD · Priority: P1-HIGH · Kingdom: Cross-kingdom
Ring: GOLD I trios-igla-race-pipeline / SR-02
Part of: #446
Blocks: SR-04 gardener, SR-ALG-03 e2e-ttt, BR-OUTPUT
Blocked by: SR-00 (issue 2), SR-01 (issue 6), SR-03 (issue 5)

Goal

The E2E TTT O(1) core. Move sampler.rs (11 K, the per-chunk sampler) + attn.rs (11 K) + gf16.rs (12 K) into SR-02. Add a versioned git dep on gHashTag/trios-trainer-igla under features = ["trios-integration"] per the SoT mandate. Do not re-implement trainer code.

Acceptance criteria

• rings/SR-02/ with I5 trinity.
• Deps: SR-00 + SR-01 + trios-trainer-igla = { git = "https://github.com/gHashTag/trios-trainer-igla", features = ["trios-integration"], rev = <pinned-sha> } + feature-gated tch (cpu-only by default).
• Public API: TrainerRunner::stream_chunks(&Job) -> impl Iterator<Item = ChunkInput>, step_one_chunk(ChunkInput) -> Result<ChunkBpb, RunnerErr> amortised O(1).
• Microbench: step_one_chunk p99 < 50 ms (CPU baseline champion.toml config).
• INV-3 GF16 closure (Lucas) check active inside step_one_chunk.
• Backward-compat re-exports in crates/trios-igla-race/src/{sampler,attn,gf16}.rs.
• Integration test: 1 fake seed, 10 chunks, BPB row written via SR-03.
• PR closes this issue, Agent: LEAD trailer.

🌻 phi² + phi⁻² = 3

[gHashTag]

➕ R14 acceptance criterion (PhD-aware)

Per EPIC #446 PhD sequencing — every Proven INV touched by this SR must have verbatim \citetheorem{} reference in the corresponding chapter. SR-02 is the E2E TTT O(1) core — critical-path issue.

Add to acceptance criteria:

• R14: Every Proven INV referenced by SR-02 (per-chunk O(1), sampler/attn/gf16 invariants) is cited verbatim via \citetheorem{} in docs/phd/chapters/*.tex, matching the Coq theorem name in crates/trios-coq/.
• R14-check: trios-doctor lint --rule R14 passes for rings/SR-02/.
• O(1) bound: per-chunk step proven O(1) per experiment_map.md (no admitted_budget consumption).
• No new Admitted. introduced (admitted_budget=5/5 LOCKED per INV-13 policy).

Ref: docs/phd/assertions/igla_assertions.json, docs/phd/experiment_map.md.

🌻 phi² + phi⁻² = 3 · Agent: LEAD

[gHashTag]

📐 R14 acceptance criteria (exact Coq names from PDF v4 / NEON ssot.chapters)

This issue's PR MUST add \citetheorem{} macros for: admit_phi_sq (Ch.8 TF3/TF9 sparse ternary MatMul) + gf16_safe_domain (Ch.9 GF vs MXFP4 ablation).

Editing rules:

• Edit lands in NEON ssot.chapters.body_md, propagated by v4/generate_from_neon.py to docs/golden-sunflowers/*.md.
• Direct edits to docs/phd/chapters/*.tex are forbidden — .tex is generated, not authored.
• Reference image (1200×800 triptych) is at assets/illustrations/ch08-tf3-tf9-ternary-matmul.png.

Constraints:

• New INV must be born Proven OR axiom-approach (🌻 EPIC — E2E TTT Pipeline O(1) · Ring-Pattern Refactor (3 GOLD × ~19 SR) #446 INV-13 policy).
• Repo admitted_budget = 5/5 LOCKED (igla-lane registry).
• Monograph ledger: 297 Qed + 141 admit/sorry + 0 Axiom — do not increase admit/sorry.

🌻 phi² + phi⁻² = 3 · Agent: LEAD

[gHashTag]

🔧 Acceptance Criteria Update — tch → candle, trios-integration → upstream-trainer

Ground truth audit before PR (R5-honest):

Fact 1 — tch не существует в trios-trainer-igla

trios-trainer-igla/Cargo.toml использует Candle (candle-core, candle-nn) с CUDA features, не tch. Acceptance criteria нужно обновить:

• ❌ feature-gated tch (cpu-only by default)
• ✅ feature gpu = ["candle-core/cuda", "candle-nn/cuda"] — зеркало upstream feature, disabled by default

Fact 2 — trios-integration feature пустая

trios-integration = []  # placeholder, все git-deps закомментированы
# Temporarily disabled due to submodule issue — use local path or fix submodule

Dep trios-trainer-igla = { git = ..., features = ["trios-integration"] } включит пустой список и не даст доступа к коду.

Предлагаемое исправление:

[features]
default = []
gpu = ["candle-core/cuda", "candle-nn/cuda"]
upstream-trainer = ["trios-trainer-igla"]

[dependencies]
trios-trainer-igla = { git = "https://github.com/gHashTag/trios-trainer-igla", rev = "09e85a0d4481b3efb998f088037a72958e3a0c65", optional = true }

Fact 3 — 85 KB live-fleet port без GPU harness = R5-unsafe

sampler.rs (11K) + attn.rs (11K) + gf16.rs (12K) — живой trainer код. Перенос без:

• champion.toml end-to-end run (требует FineWeb данные + CUDA)
• измеренного step_one_chunk p99 < 50ms (CPU baseline не определён)
• реального BPB row written via SR-03 integration test (требует Neon write)

...это hallucinated claim, не shipped ring.

---

Предлагаемый revised scope для SR-02 v0.1

Часть	R5-статус
rings/SR-02/ + I5 trinity	✅ чисто
TrainerRunner trait (абстракция) + ChunkInput/ChunkBpb типы на SR-00	✅ чисто
step_one_chunk sync API с mock implementation + BpbSink integration	✅ чисто
feature gpu = ["candle-core/cuda", "candle-nn/cuda"] disabled by default	✅ чисто
trios-trainer-igla git dep rev 09e85a0d44, gated behind upstream-trainer feature	✅ чисто
INV-3 GF16 closure check — чистая Rust-проверка	✅ чисто
Tests ≥ 10 GREEN, clippy 0 warnings	✅ чисто
85 KB port sampler/attn/gf16 + p99 < 50ms microbench	⚠️ SR-02b follow-up — требует GPU harness
BPB row written via Neon integration test	⚠️ SR-02b follow-up — требует Neon write env

SR-02 v0.1 закрывает ~60% acceptance criteria и честно фиксирует остальное как SR-02b когда GPU harness доступен через Neon.

---

Запрос: обновить acceptance criteria в issue body

Просьба к @gHashTag — подтвердить или скорректировать:

1. Заменить tch → candle в acceptance criteria
2. Заменить features = ["trios-integration"] → features = ["upstream-trainer"] (optional dep)
3. Добавить SR-02b follow-up scope для 85KB port + microbench когда GPU harness готов

После подтверждения — немедленный PR с SR-02 v0.1 skeleton.

🌻 phi² + phi⁻² = 3 · truth-first · Agent: LEAD

[gHashTag]

🔥 IGLA RACE IGNITION — cross-front status (Wave 10 pre-flight)

Bring this issue back into discussion — the operator surface has moved past paper:

What is happening on the metal (parallel army front)

• H100 SXM Pod r5axwbgsrijxho — RUNNING through trios-trainer-igla:latest (image: ghcr.io/ghashtag/trios-trainer-igla:latest)
• Env: TRIOS_SEED=10001, TRIOS_STEPS=50000, TRIOS_LR=0.04, TRIOS_HIDDEN=512, TRIOS_OPTIMIZER=muon
• 8 RunPod accounts available (runpod_accounts Neon table populated, 3 working keys verified)
• runpod-cli rewritten REST → GraphQL
• seed_agent.rs containerization fix committed locally as 429289c62 (NOT YET PUSHED to origin)

Why this issue is technically un-blocked but formally still on hold

• Acceptance criteria in this issue requires tch cpu-only by default with feature-gated — I cannot verify whether trios-trainer-igla:latest is built with tch or candle without inspecting the Dockerfile / Cargo.lock of the SoT repo (gHashTag/trios-trainer-igla).
• The H100 pod is training successfully — which means some backend works, but does not by itself satisfy AC item feat(bridge): Phase 2 — Vision + Context (GB screenshot → structured state) #2 of this issue (the backend choice and feature-gating).
• No PR exists yet that lands crates/trios-igla-race-pipeline/rings/SR-02/ — the ring scaffold itself is missing.

What is needed to actually close this issue

1. @gHashTag verify which backend is in the live trios-trainer-igla:latest image (tch or candle, cpu-only-default?).
2. Push 429289c62 to origin/fix/entrypoint-scarab-dual-mode and open PR — unblocks CI rebuild of trios-seed-agent-real:latest.
3. Open SR-02 ring scaffold PR: crates/trios-igla-race-pipeline/rings/SR-02/ with I5 trinity + trainer-runner trait wrapper + git-dep on gHashTag/trios-trainer-igla pinned by sha.
4. Validate seed semantics: current pod runs seed=10001 which is not in the gate-3 acceptance set [F_17=1597, F_18=2584, F_19=4181]. Either re-run with Fibonacci seeds, or document as warmup/calibration.

Honest scoreboard

Layer	State
SR-02 ring scaffold (this issue)	NOT shipped
H100 metal training	RUNNING (warmup seed)
seed-agent containerization	committed locally, not pushed
RunPod fleet (8 accounts)	idle, ready to fan-out after CI rebuild
WIN gate (3-seed mean BPB < 1.07063)	not measured

Not closing this issue. The metal moves but the ring is still empty. Issue stays open until SR-02 ring lands and AC checkboxes are honestly ticked.

🌻 Cross-posted from EPIC #446 Wave 10 pre-flight. Agent: LEAD. No fabricated proofs.

[gHashTag]

Requires trios-trainer-igla external repo.