Release verification

Public site: s1s.iar.dev. Source: cpaczek/s1s. The deployment is a Cloudflare Worker with SQLite Durable Object execution, using the existing account's Free plan. No account-plan upgrade or npm publication was performed.

Handoff item	Delivered behavior	Evidence
Initialize a real repository	main, isolated worktrees, public MIT repository; private data/gold excluded	Git history and tracked-file audit; no history entries for .env, data/ or private gold
Clean up and type boundaries	Native TypeScript strict/unused checks; local compiled Tailwind; validated HTTP/SSE failures; cancellation and safe file windows	Core tests, demo tests, typecheck, demo:typecheck, typecheck:ui-contracts, transport and browser smoke
Rename	System One Search / s1s package, binary, commands, docs and UI	node bin/s1s.mjs --help; CLI coverage command; published pages
Cloudflare demo	Five build-time snapshots; per-IP/global budgets, bounded queue, persistent cache and analytics	Actual Wrangler deployment, all five /api/tree responses, live find/explain and cache replay; protection tests
Polished interactive UI	Repo picker, tailored suggestions, automatic Find/Explain with override, live heatmap, provisional and final flow, cost/time, previews	Real Strapi authentication result; browser controls/error/retry/cancel/source tests
Flow transition	Matching visible file boxes move from heatmap to real graph nodes	Browser asserts real matching IDs, completed cleanup, and no animation under reduced motion
Public library/toolset	Injected-client Find/Map/Explain, portable snapshots, bounded coverage audit; Rust/Vue/GraphQL/Go/Java/Python extraction and references	src/library.ts, integration guide, multilingual/graph/snapshot tests and public benchmark rows
Landing and About	Original code-atlas mark, ink-blue/mist/lime identity, exact source-reference illustration, architecture and integration guide	Desktop/phone visual review; source import inventory for every illustrated reference
Mobile follow-up	Phone/tablet reading order, 44px controls, 16px inputs, readable graph zoom, touch pan/pinch, bounded source/results scrolling	Eight portrait/landscape interaction sizes; production pages checked at 320, 390, 768, 1024 and 1440px
No emoji icons	Local SVG/CSS controls and original SVG mark	UI source scan and rendered visual checks; semantic arrows in reference text remain notation
Retrieval benchmark follow-up	Pinned RepoQA/SWE-bench localization adapters; real grep, BM25, BM25F, local dense embeddings and repeated s1s	Results and raw rankings, corpus parity audit and strict reuse-provenance tests
Updated secret	User-authorized .env key transferred over stdin into the Worker secret store, without displaying its value	Wrangler successful secret upload, no inherited environment override, production /api/health reports live: true

Verified behavior

The live production path completed an uncached OpenCode Find request and an uncached Strapi authentication Explain request with 36 nodes and connected references. Repeating Find returned X-S1S-Cache: hit. A partial Find verdict was retained as partial rather than presented as certainty. All five repository indexes matched their catalog revisions and file counts. Private snapshot paths return 404; unknown repositories return 404; traversal and public rebuild parameters are rejected. Source previews return physical line windows.

The browser suite covers 320×568, 360×800, 390×844, 768×1024, 568×320, 800×360, 844×390 and 1024×768, plus a desktop view. It checks buttons, source previews, repo isolation, strategy override, keyboard selection, touch pan/pinch, errors, explicit retry and cancellation. Its deliberately injected HTTP 429 produces an expected console resource error; unexpected browser exceptions fail the suite. Production page checks report no unexpected console errors or horizontal page overflow. Cloudflare's injected bot-detection script uses a per-response CSP nonce; unrestricted inline script execution is not enabled.

Validation at this release: 360 core tests, 23 demo checks (12 also occur in the core suite), 6 Python benchmark tests, transport checks, strict typechecks and Chromium smoke. Hosted CI runs the offline checks, including its explicit ripgrep prerequisite. Real provider tests are separate from CI and fake-client tests; they are not inferred from a green CI badge.

What this does not prove

The engine is still bounded static retrieval. Vocabulary misses, broad-question subject selection, approximate language parsing, dynamic dependency injection, client/server boundaries and graph caps can hide relevant code. The public Rust broad-subject flow miss is preserved in the benchmark notes. No universal semantic-completeness or runtime-trace claim is made. The original handoff's research observations remain explicit in limitations.