v0.5.0 — grounding verifier, 8B base, output-language policy

Added

• Grounding verifier — a language-agnostic honesty layer for generated documents. An LLM
  enumerates every claim in a generated CV or cover letter and cites the source line that grounds
  it; a deterministic audit (documents/grounding_verifier.py) then overrides any ungrounded claim
  (token-overlap + a numeric backstop covering percentages and standalone integers — years,
  counts, team sizes) and flags coverage gaps. The SOURCE is always the BASE résumé, never the job
  description or the tailored intermediate. Cover letters route through
  CoverLetterGenerator.generate_verified() (regenerate once, keep the strictly-cleaner draft);
  tailored résumés through ResumeTailor.tailor_verified(), which surfaces the report on
  TailoredResume.grounding_report for human review (a "claims to review" panel + a key in
  tailor --json) and never auto-strips — the résumé is the document of record. Fail-safe: any
  verifier failure raises GroundingUnavailableError, never passing off an unverified document as
  clean. Replaces per-language hardcoded honesty blocklists (which were unbounded and inert on
  French) with one entailment pass; the English deterministic floor is kept as augmentation.
• Output-language policy. [llm] language = auto (mirror the job posting's language) | en |
  fr. It lives on [llm] so the cover-letter override inherits it — the generated CV and cover
  letter always resolve the same language, so one application never mixes them. French resolves
  an in-language sign-off ("Cordialement,"), a localized PDF date, and recognized French closings.
• Structured cover-letter generation. Cover letters are generated as three connected paragraphs
  via a structured (instructor) step, with deterministic honesty guards and an enforced sign-off.

Changed

• Default base model is now Qwen/Qwen3-8B-AWQ (genuine AWQ 4-bit, ~6.1 GB), replacing
  cyankiwi/Qwen3.5-4B-AWQ-4bit (kept as a pinnable fallback via JOB_APPLICATOR_LLM_MODEL /
  [llm] model). The 8B grounds stack-heavy job descriptions the 4B couldn't (measured: cover-letter
  employer-stack overclaim 5/6 → 0/5) while still fitting the 12 GB card alongside the embeddings
  (GPU_MEM=0.70, eager mode: ~8.4 GB vLLM + ~1.5 GB embeddings). Generation guardrails were
  re-tuned for the larger model, and llm.max_tokens raised to 4096.
• scripts/serve-vllm.sh puts the vLLM venv's bin on PATH so flashinfer can JIT-compile a
  kernel for a fresh model (the 8B fails with No such file or directory: 'ninja' otherwise);
  ninja now ships in the serve extra.
• Single source of truth for the litellm model id. All completion callers build it via
  utils.llm.litellm_model(config) instead of re-deriving the openai/ prefix in six places.
• Docs corrected on how the CUDA 13.0 wheel is obtained. vllm 0.23 pins torch==2.11.0,
  whose PyPI wheel is the cu13 build (bundles nvidia-*-cu13), so a plain pip install lands the
  cu13 stack with no extra index; pip selects it unconditionally and it needs a CUDA-13 driver to
  run. (README, pyproject, AGENTS.md.)

Fixed

• Serve script no longer silently piggybacks on a $PATH vLLM. scripts/serve-vllm.sh now
  uses only job-applicator's own .venv/bin/vllm or an explicit VLLM_BIN; if neither exists it
  errors with install/opt-in guidance instead of adopting whatever vllm is first on $PATH
  (historically a sibling project's), closing the last cross-project coupling in the self-host path.
• config.example.toml no longer caps max_tokens at 1024. It now ships 4096 to match the
  built-in default, so bootstrapping from the example doesn't silently truncate tailored résumés.