உங்கள் ஆர்வத்திற்கு நன்றி! இந்த திட்டத்திற்கு இரண்டு பேரம்பேசவியலாத culture விதிகள் உள்ளன — அவைதான் README-இல் உள்ள ஒவ்வொரு எண்ணையும் நம்பக்கூடியதாக மாற்றும் காரணம்.
அனைத்து production code-உம் முதலில் தோல்வியடைந்த ஒரு testலிருந்து பிறக்கிறது:
- testஐ எழுதி அது சரியான காரணத்திற்காக தோல்வியடைவதைப் பாருங்கள்.
- அதைப் pass செய்ய தேவையான குறைந்தபட்சத்தை எழுதுங்கள்.
- green ஆக இருக்கும்போதே refactor செய்யுங்கள்.
முழு bar: pnpm run typecheck && pnpm test && pnpm run build — எப்போதும்
மூன்றும் (docs link-lint மற்றும் rebrand guard pnpm testக்குள் இயங்குகின்றன
tests/docs-integrity.test.ts மூலமாக).
Geometry, atlas, billing formula, அல்லது model scopeக்கான எந்த மாற்றமும் அளவிடப்பட்ட ஒரு எண் இல்லாமல் இறங்காது. இந்த repository இந்த கட்டுப்பாட்டைச் சுற்றி கட்டமைக்கப்பட்டுள்ளது:
- Billing செலவு →
benchmarks/billing-sweep/-உடன் நிரூபியுங்கள் (count_tokensஇலவசம்; எதிர்பார்க்கப்படும் residual: பூஜ்ஜியம்). - Legibility →
benchmarks/density-frontier/-உடன் நிரூபியுங்கள் (n≥30 ஒரு armக்கு, deterministic scoring,benchmarks/*/results/-இல் commit செய்யப்பட்ட JSONL ஆதாரங்கள்). - ஒரு production defaultஐ மாற்றுவதற்கான acceptance bar: gist == text baseline மற்றும் பூஜ்ஜிய silent exact-string errors மற்றும் நேர்மறை savings.
எண்கள் இல்லாத hypothesesகள் docs/ROADMAP.md-க்கு hypothesesஆகச் செல்கின்றன — ஒருபோதும்
README-க்கு உண்மைகளாக அல்ல. இரண்டு "தெளிவான" யோசனைகள் ஏற்கனவே data மூலம் மறுக்கப்பட்டுள்ளன
(high-res பக்கம், anti-aliased atlas); இந்த செயல்முறை வேலை செய்கிறது.
pnpm install
pnpm test # full suite, ~40–90s
pnpm run dev:node # watch modeஇல் local proxyNode ≥18 (CI 20/22/24 testஆகிறது), pnpm 10 (package.jsonஇல் packageManager
மூலம் pinned).
| folder | விதி |
|---|---|
src/core/ |
runtime-agnostic (Web APIs மட்டும் — Node மற்றும் Workersஇல் இயங்குகிறது) |
src/node.ts / src/worker.ts |
host plumbing மட்டும் |
benchmarks/ |
மறுஇயக்கக்கூடிய harnesses; JSONL results ஆதாரங்கள், committed |
docs/ |
benchmarks/ (எண்கள்), architecture/ (map), ROADMAP (hypotheses), ops/ (OmniRoute) |
- Conventional commits (
feat:,fix:,perf:,docs:,refactor:,test:,chore:), relevant எண்களுடன் ஏன் என்பதை விளக்கும் body. - சிறிய, focused PRs; behavior மாற்றங்கள் அவற்றை pin செய்யும் testஉடனும், பொருந்தும்போது அவற்றை நியாயப்படுத்தும் benchmarkஉடனும் வருகின்றன.
- Client-இன்
cache_controlblocksஐ மறுஎழுதாதீர்கள், விவாதம் இல்லாமல் runtime dependencies சேர்க்காதீர்கள் (core வேண்டுமென்றே dependency-light), render pathsஇல்Math.random/timestamps பயன்படுத்தாதீர்கள் (determinism ஒரு கடினமான invariant, byte-identity மூலம் tested).