fix(fmsr): match failure-mode records on asset_class only

[ShuxinLin]

Description

Fixes two asset-class lookup bugs in the FMSR server.

Stacked on #552 (refactor/drop-fmsr-generate-tool). Merge that first; this PR's base will then switch to main.

1. Stored classes with capitals or punctuation could not be found. The CouchDB loader stores each record as _id = "fm:" + asset_class using the raw name (src/couchdb/loader.py:_make_id). The server lowercased the input and then matched exactly on fm:<key> and on asset_class == key. As a result, a record such as "asset_class": "Hydraulic Pump" was never found, and add_failure_modes created a duplicate record next to it. This stayed hidden because the only sample record is "pump".

2. Digits were stripped from class names. re.sub(r"\d+", "", ...) turned CO2 Compressor into co compressor, Type 2 Chiller into type chiller and 3-Phase Motor into phase motor.

Fix Details

• Match on asset_class only, never on _id. The server reads the records' asset_class values (a small collection, capped at 10,000) and compares the normalised names in Python.
• Simple normalisation, applied the same way to input and stored names: it ignores case, punctuation, _ and whitespace, and leaves digits alone. The caller is expected to pass a real class name. Instance ids like Pump-1 are not rewritten.
• Not-found errors suggest close matches ("Did you mean: Hydraulic Pump?") alongside the available classes, so an agent can correct its input.
• add_failure_modes keeps the stored asset_class name when it updates a record. New records are still created as fm:<normalised name>.
• Ties: if several records normalise to the same name, the lowest _id wins, deterministically.
• Agent guidance in the tool descriptions: each asset_class description now tells the agent to pass a generic equipment type rather than an asset id (for "Chiller 6", use "chiller"). get_failure_modes says to retry with a class suggested in the error. add_failure_modes says to reuse an existing class name where one fits, to avoid near-duplicate classes.

Impact on Benchmarking

• No change to baselines
• Baseline change (possible)
  • Class-name inputs such as pump, Pump and hydraulic_pump behave as before for the bundled pump sample.
  • Inputs that are really instance ids (Pump-1) no longer map to pump. They now return a not-found error that suggests pump.
  • No before/after benchmark run was done.

Related Issues

• Found while reviewing the FMSR server after refactor(fmsr): drop LLM generation, fix asset_class matching #552.

Verification Steps

1. uv run pytest src/servers/fmsr -q passes (32 tests).
2. 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.

New tests cover:

• stored names with capitals or punctuation, and _ids not in the fm:<class> form
• digits kept in class names
• instance ids rejected, with a suggestion
• _id never used for matching
• duplicate normalised classes
• add updating the matched record in place

Test gaps closed:

• FakeDatabase.find returned the stored dicts themselves, so write tests passed even if save was never called. It now returns copies.
• FakeDatabase now enforces _rev conflicts the way CouchDB does.
• There are now tests for a failed save and for a concurrent update (conflict).

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting rules. ruff check still reports the file's existing List/Optional typing-style warnings (UP006/UP045).
• I have signed off my commits (DCO).