Add local-model variant of the opportunities agent

[makegov-mark]

Summary

• New examples/opportunities-agent/local_agent.py — the mirror of agent.py swapped onto any OpenAI-compatible endpoint (Ollama, LM Studio, vLLM, llama.cpp). Same tool catalog, same loop shape; the two files diff cleanly so the diff itself is the lesson on where the Anthropic and OpenAI tool-use protocols differ.
• Hardened LOCAL_SYSTEM_PROMPT + tighter MAX_TURNS=6 — soft guidance like "two empty searches is a signal to stop" that Sonnet self-regulates on lets a ~30B local model burn every available turn keyword-grinding. Numbered hard rules and per-phase call caps turn a failed local run into a working demo.
• sam_url now comes from Tango (in both agent.py and local_agent.py) — the brief includes a clickable SAM.gov link. Tango computes it using the latest notice id with hyphens stripped, which is what SAM.gov actually accepts; the original opportunity_id (hyphenated) silently 404s on SAM. Done by requesting sam_url via the shape= parameter rather than constructing the URL client-side.
• As a side effect of correcting the field names in the new shape= string (naics_code/psc_code, not naics/psc), six fields the original _trim_opportunity was aspirationally requesting actually show up in the tool result for the first time.
• just local-agent, openai>=1.50 in the examples dependency group, and LOCAL_MODEL/LOCAL_BASE_URL/LOCAL_API_KEY documented in .env.example.

Test plan

• just agent (default question) — still works against Anthropic, sam_url now appears in the brief
• just local-agent against LM Studio (google/gemma-4-26b-a4b-qat) — clean 3-call run, brief includes correct SAM.gov link (hyphens stripped, latest notice id)
• just local-agent "<harder SDVOSB question>" against remote vLLM endpoint — completes under MAX_TURNS, honest "no incumbent found"
• Confirmed generated SAM URLs return 200 (curl -I https://sam.gov/opp/<id>/view)

Caveats

• Local-model tool-calling on smaller models is stochastic. LM Studio's chat template for some Gemma builds leaks gpt-oss-style `<|channel|>` markers into the assistant content, which produces empty `tool_calls` and stalls the loop. Worked around at the prompt level, but ultimately a template-side issue you may want to fix in LM Studio before demoing.
• This example is not run in CI (paid LLM API + non-deterministic output).