[AIENG-601] Client-Side Embedding generation poc implementation

[sandeepbarnwal]

Caution

This is only a POC implementation and should not be merged.

AIENG-601: BYOLLM — Client-Side Embeddings via Customer's LLM Endpoint

Summary

Implements the CLI side of the Bring Your Own LLM (BYOLLM) feature. When a workspace is configured for embeddingMode=client, the CLI fetches embedding configuration
from the server options response, calls the customer's OpenAI-compatible LLM endpoint to generate file embeddings, and uploads the resulting vectors to the
Launchable server. Both record commit and record build support this path.

---

Python CLI

• commands/record/commit.py — reads embeddingMode, embeddingEndpoint, embeddingModel, embeddingDimensions, embeddingAugmentation, and embeddingProvider from the
  commits/collect/options response. Env vars (SMART_TESTS_EMBEDDING_ENDPOINT, SMART_TESTS_EMBEDDING_MODEL) override server-provided values. Validates endpoint and API
  key before enabling embeddings; degrades gracefully with a warning if either is missing. record build inherits this path automatically via commit.callback().
• utils/env_keys.py — added SMART_TESTS_EMBEDDING_ENDPOINT, SMART_TESTS_EMBEDDING_API_KEY, SMART_TESTS_EMBEDDING_MODEL constants.

---

Java — new embedding/ package

• RemoteEmbeddingStrategy — calls OpenAI-compatible endpoint; batches up to 1900 files / 210K tokens; rate-limits; retries on 429/5xx with exponential backoff;
  L2-normalizes vectors; skips blank files.
• AugmentedEmbeddingStrategy — prepends server-generated file summaries to content before embedding.
• EmbeddingStrategyFactory — builds the right strategy; takes two separate CloseableHttpClient instances — one plain (LLM endpoint), one Launchable-auth-configured
  (summaries endpoint) — to prevent auth header leakage.
• TokenizerFactory — selects tokenizer from server-provided provider (openai/azure_openai → Cl100kTokenizer, custom → NoopTokenizer); falls back to URL-host sniffing
  when provider is null.
• Supporting types: EmbeddingStrategy (interface), Cl100kTokenizer, NoopTokenizer, ServerRepoContextProvider, FileToEmbed, FileEmbeddingResult, Summaries.

Java — modified

• commits/Main.java — added CLI options -embedding-endpoint, -embedding-model, -embedding-dimensions, -embedding-augmentation, -embedding-provider. Builds two
  separate HTTP clients. API key is read via System.getenv() only — never passed as a CLI arg to avoid ps aux exposure.
• commits/CommitGraphCollector.java — setEmbeddingStrategy(), transferWithEmbeddings(), file consumer integration.

---

Tests

• AugmentedEmbeddingStrategyTest — prepend format, empty summary, metadata preservation, delegate modelName/dimensions
• ServerRepoContextProviderTest — MockServer: returns summaries, correct request body, non-2xx throws IOException
• EmbeddingStrategyFactoryTest — no-augmentation, augmentation wrapping, end-to-end embed, provider-based tokenizer selection
• RemoteEmbeddingStrategyTest — added provider-based tokenizer tests (openai, azure_openai, custom, null fallback to URL sniffing)
• EmbeddingAllTests — JUnit4 suite covering all embedding tests under one Bazel target

---

Security

• SMART_TESTS_EMBEDDING_API_KEY is read via System.getenv() in the JVM only, never passed as a CLI argument — prevents exposure via ps aux.
• Two separate HTTP clients ensure the Launchable workspace token is never sent to the customer's LLM endpoint and vice versa.

---

Test Plan

• bazelisk test //src/test/java/com/launchableinc/ingest/embedding:EmbeddingTests — 24 tests pass
• record commit --source . end-to-end — 22 files embedded via OpenAI, tokenizer selected from provider=openai without any env var override
• record build --build --source . end-to-end — same embedding path, build recorded successfully
• Missing endpoint → graceful warning, embeddings skipped, commit/build still recorded
• Missing API key → graceful warning, embeddings skipped, commit/build still recorded