Repository hygiene — stale PR, branch model, misleading closed issues, doc drift

[lucaosti]

Context

Housekeeping found during the project review that does not belong to a feature milestone.

Tasks

Stale pull request

• PR feat: implement study actions #122 ("feat: implement study actions") has been open since 11 June and targets mvp, not master. Its content is part of the mvp-testing integration (Integrate the mvp-testing development line into master #130). Once that work lands on master, close feat: implement study actions #122 with a comment pointing at the integration, or retarget it if it is still the vehicle for the merge.

Branch hygiene

• mvp and mvp-testing are kept deliberately — MVP testing work continues there. But mvp is currently 5 commits behind mvp-testing and both will diverge from master further during integration. Agree and document the branch model in AGENTS.md: what each branch is for, what merges where, and when they are reconciled.

Issue accuracy

• RAG · Upgrade embedding model (Qwen3-Embedding 0.6B o bge-m3) #84 ("Upgrade embedding model") is closed as done, but the default embedding model is still GTE-Large FP16 (workers/qvac-service/src/models.js). What shipped was configurability via QVAC_EMB_SRC, not the upgrade. Add a clarifying comment; the actual upgrade remains open as Embedding model upgrade — GTE-Large FP16 → Qwen3-Embedding-0.6B #115.
• RAG · Sostituire pymupdf4llm con Docling per parsing PDF #85 ("Sostituire pymupdf4llm con Docling") is closed as done, but Docling is an opt-in parser (USE_DOCLING=false by default) with fallback, not a replacement. Add a clarifying comment.
• RAG · RAGAS evaluation framework (scripts/eval_rag.py) #83 references scripts/eval_rag.py, which does not exist. Clarify what was delivered — see Consolidate the four overlapping RAG evaluation artefacts onto one baseline #141.

Documentation drift guards

• The README's RAG variable table drifted from the code (SKIP_CHROMA_INDEX was documented as true, code defaults to false). Consider a test or CI check asserting documented defaults match Settings, so this class of drift cannot recur silently.

Acceptance criteria

• No pull request open against a branch that is not part of the agreed model
• The branch model documented in AGENTS.md
• Every misleading closed issue carries a comment stating what was actually delivered
• Documented configuration defaults match code, ideally enforced automatically

Why this matters

Per AGENTS.md § Keeping GitHub honest: a Done that does not match the code is worse than an open issue, because it stops anyone from looking.