feat: read repo list and model name from config files, defaulting to prior hardcoded behavior

[man4ish]

What

Wires two existing-but-unused config files into the code, with the current hardcoded behavior kept as the fallback in both cases. Designed to be functionally identical to today's behavior — nothing changes unless someone explicitly edits a config file.

1. scripts/build_index.py now reads the repo list from configs/repos.yaml (a repos: list of names) if the file exists, parses, and is non-empty. The old hardcoded 19-repo list is kept in place, unchanged, as HARDCODED_REPO_NAMES — used whenever the YAML is missing, empty, or fails to parse.
2. requirements.txt: added PyYAML==6.0.2, pinned to the version actually installed and working in this environment (checked via pip freeze, same pinning approach as PR chore: pin requirements.txt to known-working versions #9 — not guessed).
3. rag/engine.py: the hardcoded "llama3" at ollama_generate()'s default arg (was line 83) and in stream_llm()'s payload dict (was line 274) now both read from a single LLM_MODEL module-level constant, resolved once at import via _load_llm_model(), which reads llm_model: from configs/index_config.yaml. If the file is missing, empty, fails to parse, or doesn't set llm_model, it falls back to "llama3" — same pattern already used for OLLAMA_URL via os.environ.get(..., default).
4. configs/index_config.yaml (previously empty) now contains llm_model: llama3 — i.e. today's value, made explicit and editable.

This is designed to be a no-op today

Both configs/repos.yaml and configs/index_config.yaml already existed in the repo (the former already listed the same 19 repos in the same order; the latter was empty). This PR doesn't change their content in a way that alters behavior — it just makes the code actually read them, with fallbacks that reproduce the prior hardcoded behavior exactly.

Repo-list before/after comparison

Ran build_index.build_index() (with load_documents monkey-patched to just record repo paths instead of doing real I/O/embedding) twice:

• With configs/repos.yaml present (its current, unmodified form): 19 repos resolved, in order.
• With configs/repos.yaml temporarily renamed away (forcing the hardcoded fallback — confirmed by the ⚠️ Could not read ... falling back to hardcoded repo list warning firing): 19 repos resolved, in order.

diff between the two captured lists: identical, byte-for-byte, all 19 repos, same order. The YAML was restored to its original location/content immediately after the test (git status confirms no stray changes).

pytest results vs. 174-test baseline

• Baseline (this session, prior to any change): 174 passed
• After this change: 174 passed

Identical — no new failures, no new skips.

Manual /rag/query test

1. Default config (llm_model: llama3, i.e. unset/prior behavior): started the API for real (existing FAISS index, live Ollama), POSTed a real query:

POST /rag/query {"query": "What model does the streaming endpoint use for generation?"}
→ HTTP 200, full grounded answer + 5 sources, context_used: 5

Confirms nothing broke relative to the pre-change baseline.

2. Override test: temporarily edited configs/index_config.yaml to llm_model: llama3.1:8b-instruct-q4_K_M (a different model actually loaded in this Ollama instance), confirmed rag.engine.LLM_MODEL picked it up at import (llama3.1:8b-instruct-q4_K_M), restarted the API, and queried:

POST /rag/query {"query": "Say the word BANANA and nothing else."}
→ HTTP 200, answer: "BANANA."

No crash, and the terse instruct-tuned response (vs. the more verbose base-model answer in test 1) confirms the override model was actually invoked, not silently ignored. Config reverted back to llm_model: llama3 before committing — the committed value reproduces prior behavior exactly.

Scope

4 files touched: configs/index_config.yaml, rag/engine.py, requirements.txt, scripts/build_index.py. No other files changed.

⚠️ Do not auto-merge

Per instructions, this PR is not to be merged automatically — same as PR #9. Results reported above for review; merge only after explicit confirmation.