fix(providers): CursorAgent deliver prompt via @path to saved file

[Spencer1O1]

Summary

99 saves the full prompt to tmp/99-*-prompt before make_request. Stock CursorAgentProvider passes the entire multiline XML query as the --print argv argument, which breaks on Windows when Neovim cwd differs from the project root.

Deliver the prompt via a single-line --print message that references the saved file with @absolute/path instead.

Same pattern as #154 proposes for OpenCode (--file).

Closes #180

Verification

Windows 11 — search in C:\Dev\test. Before fix: multiline <Context> on --print → agent "Your message looks incomplete". After:

{"msg":"saved prompt to file","path":"C:\\Dev\\test\\tmp/99-6155-prompt"}
{"msg":"make_request","command":["cursor-agent","--trust","--force","--model","composer-2.5-fast","--print","Read and follow every instruction in @C:/Dev/test/tmp/99-6155-prompt using your file tools, then complete the task exactly as specified in that file."]}
{"msg":"stdout","data":"Task complete. Results written to `C:/Dev/test/tmp/99-6155`.\n"}

Linux (WSL) — search, same branch:

{"msg":"saved prompt to file","path":"/home/spencerls/tmp/99-9659-prompt"}
{"msg":"make_request","command":["cursor-agent","--trust","--force","--model","composer-2.5-fast","--print","Read and follow every instruction in @/home/spencerls/tmp/99-9659-prompt using your file tools, then complete the task exactly as specified in that file."]}
{"msg":"stdout","data":"Task complete. Results are in `/home/spencerls/tmp/99-9659`.\n"}

Test plan

• Provider spec: print arg is single-line, contains @ and 99-*-prompt, no <Context>
• Manual Windows: search — agent reads prompt file and runs task
• Manual Linux: search — same behavior

[cursor]

Bugbot is not enabled for this team, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[cursor]

Bugbot is not enabled for this team, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[Spencer1O1]

Re: Sentry — valid edge case, fixed in 44e14c2. The cwd fallback was returned without a filereadable check. Now both candidates are tried in order and only a readable path is returned. Normal flow is unaffected: finalize() creates the prompt file via _ready_request_files before save_prompt/make_request. If no candidate is readable (I/O failure), we still fall back to the primary normalized path so the agent error is visible rather than silently using a wrong cwd path.