v3.10.17

## What ships

**Pretrain self-learning from a repo's GitHub history** (ADR-077). One script,
zero config, ~3 ms per trajectory. Turns the day-one "0 patterns, 0
trajectories" problem into a one-liner.

```bash
node v3/@claude-flow/cli/scripts/pretrain-from-github.mjs
# → 80 trajectories trained from 50 commits + 30 issues
# → +95 trajectoriesRecorded, +85 patternsLearned, +80 neuralPatternCount
# → 100% retrieval match rate across 10 sample queries
```

## Why

ADR-074–076 (3.10.14–3.10.16) fixed *honesty*, *coherence*, and *retrieval
quality*. But a fresh ruflo install still started empty — every "did learning
happen?" call legitimately returned 0 until many real sessions had run. This
release closes that gap by seeding from a signal source every repo already has:
its own commits and issues.

Each item flows through the same code paths real-time learning uses (no
shortcuts) — distillAndSerialise (ADR-076) → recordTrajectory (ADR-074) →
neural store seed. That also closes the ADR-075 consistency note
"globalStats moved but neural_patterns stayed empty" by writing to both
stores from the same script.

## Measured proof

| | Before | After | Δ |
|---|---:|---:|---:|
| trajectoriesRecorded | 0 | 95 | +95 |
| patternsLearned | 0 | 85 | +85 |
| neuralPatternCount | 15 | 95 | +80 |
| Trained / harvested | — | — | **80/80** |
| Avg pretrain latency | — | — | **3.36 ms/item** |
| Retrieval match rate (N=95) | — | — | **100% (10/10)** |
| Avg query latency | — | — | **7.67 ms** |

Run JSONs:
- `docs/benchmarks/runs/pretrain-from-github-latest.json`
- `docs/benchmarks/runs/pretrained-retrieval-latest.json`

## What's in the box

- `scripts/pretrain-from-github.mjs` — env-configurable harvester
  (`COMMITS`, `ISSUES`, `SOURCE`, `BENCH_JSON`)
- `scripts/benchmark-pretrained-retrieval.mjs` — after-pretrain validator
- `__tests__/pretrain-from-github.test.ts` — CI guard with embedded fixture
  (no live git/gh in tests; auto-picked-up by v3-ci.yml)
- `v3/docs/adr/ADR-077-pretrain-from-history.md`
- `v3/docs/learning/self-learning-usage.md` — copy-paste guide covering all
  three learning paths plus pretrain

## Reproduce

```bash
git clone https://github.com/ruvnet/ruflo && cd ruflo
npm install && ( cd v3/@claude-flow/cli && npx tsc -b )
node v3/@claude-flow/cli/scripts/pretrain-from-github.mjs
node v3/@claude-flow/cli/scripts/benchmark-pretrained-retrieval.mjs
( cd v3/@claude-flow/cli && npx vitest run __tests__/pretrain-from-github.test.ts )
```

## Honest limits

- Standalone-process drift: when the script runs outside the live MCP daemon,
  `sonaCoordinator` and `memory-bridge` start empty. The script's consistency
  block flags this explicitly. From inside the daemon both stores are warm.
- 100% match-rate ≠ semantic relevance. Pretrain proves the wiring; ADR-076's
  MRR benchmark is the right gauge for relevance quality.
- Commits are all recorded as `success` (no outcome signal). A "was this commit
  reverted?" verdict refinement is tracked for follow-up.

## Install

```bash
npx ruflo@3.10.17                 # or @latest
npx ruflo@alpha                   # legacy compat
npx ruflo@v3alpha                 # legacy compat
```

All three packages (`@claude-flow/cli`, `claude-flow`, `ruflo`) are at
3.10.17 on `latest`, `alpha`, and `v3alpha`.