refactor(fmsr): drop LLM generation, fix asset_class matching

[ShuxinLin]

Description

Simplifies the FMSR MCP server so it only stores and looks up failure-mode catalogs. It no longer calls an LLM of its own. This PR also fixes how asset classes are matched.

1. Drop generate_failure_modes. The tool called a separate LLM inside the server (FMSR_MODEL_ID, default WatsonX Llama-3.3-70B). Because of that:

• fmsr results depended on that fixed model instead of the agent being tested;
• without WatsonX credentials the tool quietly returned {"error": "LLM unavailable"}, so lite-profile fmsr scenarios scored badly and no error was shown;
• the server needed its own LLM credentials and billing.

Agents now come up with failure modes from their own domain knowledge and save them with add_failure_modes. The failure-mode/sensor mapping already works this way. MCP sampling was considered, but none of the three runners support it yet.

2. Fix asset-class matching (merged in from #553).

• Stored names with capitals or punctuation could not be found. The loader stores records as fm:<raw asset_class>, while the server lowercased the input and looked up fm:<key>. Records such as "Hydraulic Pump" were never found, and add_failure_modes created duplicates of them.
• Digits were stripped from class names. CO2 Compressor became co compressor, and 3-Phase Motor became phase motor.

Changes

• Documentation / Tutorial update
• Refactoring (no logic change): not quite; this removes a tool and changes matching

Server (src/servers/fmsr/main.py)

• Removed: generate_failure_modes, its result model, the prompts and the LLM setup (FMSR_MODEL_ID). The server now makes no LLM calls.
• Matching: records are matched on asset_class only, never _id. Input and stored names go through the same normalisation, which ignores case, punctuation, _ and whitespace and leaves digits alone.
• Ties: if several records share the same normalised name, the lowest _id wins.
• Errors: a not-found error suggests close matches ("Did you mean: Hydraulic Pump?") and lists the available classes.
• Updates: add_failure_modes keeps the stored asset_class name. New records are created as fm:<normalised name>.
• Tool descriptions: they tell agents to pass a generic equipment type, not an asset id ("Chiller 6" → "chiller"), to retry with a suggested class, and to reuse existing class names.
• Server instructions: they now say the agent generates failure modes and mappings itself and saves them with add_failure_modes.

Tests (src/servers/fmsr/tests/)

• Removed: the generate-tool tests and fixtures. The registration test checks that the tool set is exactly {get_failure_modes, add_failure_modes}.
• Test database stand-in: FakeDatabase.find now returns copies. It used to return the stored dicts themselves, so write tests passed even if save never ran. It also rejects conflicting writes (_rev) the way CouchDB does.
• New tests:
  • stored names with capitals or punctuation, and _ids not in the fm: form
  • digits kept in class names
  • instance ids rejected, with a suggestion
  • _id never used for matching
  • duplicate normalised classes
  • updates applied to the matching record
  • a failed save
  • a conflicting concurrent update

Docs and scripts

• README: tool table.
• docs/mcp-servers.md: tool list and requirements.
• docs/running_benchmark.md: removed the FMSR_MODEL_ID row, the "WatsonX — needed more often than it looks" section and the related troubleshooting row.
• docs/opencode-agent.md: removed the generate smoke tests and renumbered the rest.
• scripts/preflight_run.sh: removed the WatsonX check.

Impact on benchmarking

• Credentials: fmsr scenarios no longer need WatsonX credentials.
• Class-name input: pump, Pump and hydraulic_pump give the same results as before for the bundled pump sample.
• Instance ids: inputs such as Pump-1 no longer map to pump. They now get a not-found error that suggests pump.
• Not run: no before/after benchmark comparison. The scenario questions aren't in this repo, so I couldn't check whether any of them tell the agent to call generate_failure_modes. Local reports/ for 902–916 show answers taken from the catalog.

Testing

• uv run pytest src/servers/fmsr -q passes (32 tests).
• uv run pytest src/ -k "not integration": 19 failures, the same set that fails on main (iot test_invalid_site, test_static_json_scorer, test_file_exporter). None are in fmsr.
• ruff format: clean. ruff check still reports the file's existing List/Optional typing-style warnings.

Not in this PR

• notebook/kdd_tutorial/03_mcp_fmsr_fixed.ipynb still references generate_failure_modes. It is untracked locally.

Checklist

• I have signed off my commits (DCO).