trust cursor workspaces in headless runs

[Svector-anu]

what

always pass cursor's headless --trust flag, while keeping --force limited to write-mode runs. add and register a focused adapter regression test, then update the generated capability metadata and harness docs to match the real invocation.

why

read-only skills never received --force, and cursor starts every ci run with a fresh temporary home. the workspace therefore had no saved trust decision, so cursor exited before executing the prompt with pass --trust, --yolo, or -f if you trust this directory.

reproduced in fork runs 33149126007 and 33149186413.

fix

• pass --trust for both read-only and write runs
• retain --force only outside read-only mode
• test the exact argument split with a stub cursor agent
• register the test in the hand-listed ci workflow
• update the invocation metadata, generated manifest, and both harness docs

verification

mutation with only --trust removed:

rc=1
read-only cursor run omitted --trust

restored:

cursor adapter trust tests passed

real fork dispatch: run 33187785314

• harness: cursor | auth: native-key | model: gpt-5.1
• capability mode: read-only
• completed successfully in 4m24s
• captured 4,239 bytes of narrative output
• scorer recorded score 4
• history appended harness: cursor, model: gpt-5.1
• no --force was supplied by the tested adapter path

the skill used its documented internal-log fallback rather than x.ai in this cursor run; that limitation is stated in the captured output and is not hidden as an api-backed result.

local checks:

• cursor adapter regression: passed
• harness manifest tests: passed
• config validation: clean
• git diff --check: clean
• local shellcheck unavailable; real ci is the source of truth for that gate.