docs: add orchestration dispatch queue and research findings

TODO.md tracks tasks dispatched to satellite repos (go-i18n phases 1-3).
FINDINGS.md records go-i18n architecture assessment and CoreDeno PR #9 review.
Phase 2 expanded with 1B classification pipeline based on LEK benchmarks.

Co-Authored-By: Virgil <virgil@lethean.io>

FINDINGS.md

@@ -0,0 +1,77 @@
+# FINDINGS.md — Core Go Research
+
+## go-i18n (forge.lthn.ai/core/go-i18n)
+
+**Explored**: 2026-02-19
+**Location**: `/Users/snider/Code/host-uk/go-i18n`
+**Module**: `forge.lthn.ai/core/go-i18n`
+**State**: 20 commits on main, clean, all tests pass
+**Lines**: ~5,800 across 32 files (14 test files)
+**Deps**: only `golang.org/x/text`
+
+### What It Is
+
+A **grammar engine** — not a translation file manager. Provides:
+
+1. **Forward composition**: `PastTense()`, `Gerund()`, `Pluralize()`, `Article()`, handlers
+2. **Reverse grammar**: Tokeniser reads grammar tables backwards to extract structure
+3. **GrammarImprint**: Feature vector projection (content → grammar fingerprint, lossy)
+4. **Multiplier**: Deterministic training data augmentation (no LLM)
+
+Consumers (core/cli, apps) bring their own translation files. go-i18n provides the grammar primitives.
+
+### Current Capabilities
+
+| Feature | Status | Notes |
+|---------|--------|-------|
+| Grammar primitives (past/gerund/plural/article) | Working | 100 irregular verbs, 40 irregular nouns |
+| Magic namespace handlers (i18n.label/progress/count/done/fail/numeric) | Working | 6 handler types |
+| Service + message lookup | Working | Thread-safe, fallback chain |
+| Subject builder (S()) | Working | Fluent API with count/gender/location/formality |
+| Plural categories (CLDR) | Working | 7+ languages |
+| RTL/LTR detection | Working | 12+ RTL languages |
+| Number formatting | Working | Locale-specific separators |
+| Reversal tokeniser | Working | 3-tier: JSON → irregular → regular morphology |
+| GrammarImprint similarity | Working | Weighted cosine (verbs 30%, tense 20%, nouns 25%) |
+| Multiplier expand | Working | Tense + number flipping, dedup, round-trip verify |
+
+### What's Missing / Incomplete
+
+| Gap | Priority | Notes |
+|-----|----------|-------|
+| Reference distribution builder | High | Process scored seeds → calibrate imprints |
+| Non-English grammar tables | Medium | Only en.json exists, reversal needs gram.* per language |
+| Ambiguity resolution | Medium | "run", "file", "test" are both verb and noun |
+| Domain vocabulary expansion | Low | 150+ words, needs legal/medical/financial |
+| Poindexter integration | Deferred | Awaiting Poindexter library |
+| TIM container image | Deferred | Distroless Go binary for confidential compute |
+
+### Key Architecture Decisions
+
+- **Bijective grammar tables**: Forward and reverse use same JSON → reversal is deterministic
+- **Lossy projection**: GrammarImprint intentionally loses content, preserves only structure
+- **No LLM dependency**: Multiplier generates variants purely from morphological rules
+- **Consumer translations are external**: go-i18n doesn't ship or manage app-specific locale files
+- **gram.* keys are sacred**: Agents MUST NOT flatten — grammar engine depends on nested structure
+
+### pkg/i18n in core/go
+
+- Full i18n framework with 34 locale files — but locale data is bad/stale
+- Only imported by `pkg/cli/` which has been extracted to `core/cli`
+- Effectively orphaned in core/go
+- Can be removed once core/cli imports go-i18n directly
+- The locale files need full rework, not migration
+
+---
+
+## CoreDeno (PR #9 — merged)
+
+**Explored**: 2026-02-19
+
+Deno sidecar for core-gui JS runtime. Go↔Deno bidirectional bridge:
+- Go→Deno: JSON-RPC over Unix socket (module lifecycle)
+- Deno→Go: gRPC over Unix socket (file I/O, store, manifest)
+- Each module in isolated Deno Worker with declared permissions
+- Marketplace: git clone + ed25519 manifest verification + SQLite registry
+
+10 security/correctness issues found and fixed in review.

TODO.md

@@ -0,0 +1,52 @@
+# TODO.md — Core Go Dispatch Queue
+
+Tasks dispatched from core/go orchestration to satellite repos.
+Format: `- [ ] REPO: task description` / `- [x]` when done.
+
+---
+
+## go-i18n (forge.lthn.ai/core/go-i18n)
+
+### Phase 1: Harden the Engine
+
+- [ ] **go-i18n: Add CLAUDE.md** — Document the grammar engine contract: what it is (grammar primitives + reversal), what it isn't (translation file manager). Include build/test commands, the gram.* sacred rule, and the agent-flattening prohibition.
+- [ ] **go-i18n: Ambiguity resolution for dual-class words** — Words like "run", "file", "test", "check", "build" are both verb and noun. Tokeniser currently picks first match. Need context-aware disambiguation (look at surrounding tokens: article before → noun, after subject → verb).
+- [ ] **go-i18n: Extend irregular verb coverage** — Audit against common dev/ops vocabulary. Missing forms cause silent fallback to regular rules which may produce wrong output (e.g. "builded" instead of "built").
+- [ ] **go-i18n: Add benchmarks** — `grammar_test.go` and `reversal/tokeniser_test.go` need `Benchmark*` functions. The engine will run in hot paths (TIM, Poindexter) — need baseline numbers.
+
+### Phase 2: Reference Distribution + 1B Classification Pipeline
+
+#### 2a: 1B Pre-Classification (based on LEK-1B benchmarks)
+
+- [ ] **go-i18n: Classification benchmark suite** — `classify_bench_test.go` with 200+ domain-tagged sentences. Categories: {technical, creative, ethical, casual}. Ground truth for calibrating 1B pre-tags.
+- [ ] **go-i18n: 1B pre-sort pipeline tool** — CLI/func that reads JSONL corpus, classifies via LEK-Gemma3-1B, writes back with `domain_1b` field. Target: ~5K sentences/sec on M3.
+- [ ] **go-i18n: 1B vs 27B calibration check** — Sample 500 sentences, classify with both, measure agreement. 75% baseline from benchmarks, technical↔creative is known weak spot.
+- [ ] **go-i18n: Article/irregular validator** — Lightweight funcs using 1B's strong article (100%) and irregular base form (100%) accuracy as fast validators.
+
+#### 2b: Reference Distributions
+
+- [ ] **go-i18n: Reference distribution builder** — Process 88K scored seeds through tokeniser + imprint. Pre-sort by `domain_1b` tag. Output: per-category reference distributions as JSON.
+- [ ] **go-i18n: Imprint comparator** — Distance metrics (cosine, KL divergence, Mahalanobis) against reference distributions. Classification signal with confidence. Poindexter integration point.
+- [ ] **go-i18n: Cross-domain anomaly detection** — Flag texts where 1B domain tag disagrees with imprint classification. Training signal or genuine cross-domain text — both valuable.
+
+### Phase 3: Multi-Language
+
+- [ ] **go-i18n: Grammar table format spec** — Document the exact JSON schema for `gram.*` keys so new languages can be added. Currently only inferred from `en.json`.
+- [ ] **go-i18n: French grammar tables** — First non-English language. French has gendered nouns, complex verb conjugation, elision rules. Good stress test for the grammar engine's language-agnostic design.
+
+---
+
+## core/go (this repo)
+
+- [ ] **core/go: Remove pkg/i18n dependency** — Once core/cli imports go-i18n directly, remove `pkg/i18n/` from this repo. The locale files are bad data and shouldn't be migrated.
+- [ ] **core/go: Update go.work** — Add go-i18n to the workspace for local development (`go work use ../go-i18n`).
+
+---
+
+## Workflow
+
+1. Virgil (this session) writes tasks above after research
+2. Second GoLand session opens the target repo and works from this TODO
+3. When a task is done, mark `[x]` and note the commit/PR
+4. If a task reveals new work, add it here
+5. Scale to other repos once pattern is proven