[feat] Add fit-aware Hugging Face model search

[willsarg]

Problem

ara models search currently returns Hub popularity metadata but cannot help a user understand whether a remote model is plausibly compatible with a selected ARA engine or likely to fit this machine.

Users must leave ARA, inspect repository files and formats, estimate download size, and mentally compare that with profile before deciding what to characterize.

Proposed change

Add an opt-in fit-aware mode to ara models search, with an explicit engine target, for example:

ara models search "small instruct" --fits --engine cpu

For the bounded search result set, enrich each candidate where authoritative metadata is available with:

• target engine and format compatibility
• estimated download/weight size
• analytic fit status: likely_fits, does_not_fit, or unknown
• estimated usable context when the architecture facts support it
• a concise reason and provenance for unknown results

Keep ordinary search fast and unchanged when the option is absent.

Acceptance criteria

• Default search output and request behavior remain backward-compatible.
• Fit-aware mode never downloads weights, installs or imports an engine, loads a model, or mutates the local cache/database.
• It reuses profile/estimate budget math rather than introducing a separate fit formula.
• Ambiguous repositories with multiple candidate artifacts or incomplete size/architecture metadata are labeled unknown; ARA does not select a quant or revision silently.
• Results remain analytic discovery hints and cannot authorize run, serve, or benchmark.
• Fit status is not presented as a quality score or research evaluation.
• Per-result metadata failures degrade to typed unknown results without discarding otherwise valid search results.
• Human and JSON outputs identify the engine, evidence source, size unit, and analytic—not measured—status.
• Tests cover compatible fit, incompatible format, too large, ambiguous artifacts, missing metadata, offline/partial failure, and unchanged ordinary search.

Non-goals

• Automatically downloading or characterizing a result.
• Quality ranking, benchmark substitution, or license-policy enforcement.
• Claiming an exact safe ceiling from remote metadata.

Independence

This is a read-only enrichment of Hub search. It does not depend on all-engine cached recommendations, #49, #56, or any fleet/provenance work.