[xavierforge]

Closes #227

Current behavior

The prompt shaping a session lives only in memory, on ContextFiles::current_prompt_name. It is set at startup from default_prompt or --load-prompt, and updated at runtime through apply_prompt_mode plus a few direct assignment sites. It surfaces in exactly one place, the status bar. The session file has no field for it, so a saved session cannot say whether it ran under code, plan, review, or a user prompt file, even though a prompt's %%mode= directive now drives permission enforcement.

What changes

• Session gains #[serde(default)] pub prompt: Option<PromptRef>, where PromptRef is { name, source } and source is BuiltIn or UserFile.
• context::prompts gains source_of(name), a small helper that checks the same override directories load() already walks (global prompts dir, the project's data/prompts/, .zerostack/prompts/) to classify a resolved prompt's origin, without changing load()'s HashMap<String, String> return shape.
• Startup prompt resolution writes the field in both the default_prompt block and the --load-prompt block.
• apply_prompt_mode now takes the session and writes the field, so its five call sites (/prompt <name>, /review, chain phase transitions, both dot-prefix branches) pick it up by construction. The three sites that assign current_prompt_name directly instead of going through apply_prompt_mode (/prompt default, and the two dot-prefix restore paths) get the same treatment.
• No switch history is kept; the last prompt active wins, matching what the status bar already shows.
• #[serde(default)] keeps existing session files loading with prompt: None.

Tests

• src/session/mod.rs schema plus a save/load round trip test and a test that a session file written before this field existed still loads, with prompt: None.
• src/context/prompts.rs: source_of unit tests for a built-in prompt, a user file shadowing a built-in name, and a prompt that only exists as a user file.
• src/tests/prompt_mode_tests.rs: apply_prompt_mode now asserts the session's recorded prompt name and source, including that an unknown prompt name is a no-op that leaves the previous record untouched, and that switching twice leaves the record reflecting the prompt active last.

Verified with cargo fmt --check and cargo clippy --all-targets --all-features -- -D warnings, both clean.

[gi-dellav]

Please fix conflicts; for the rest, it's ready to be merged

[xavierforge]

Conflicts fixed (rebased onto current main; the only overlap was Session gaining fields from both #230 and this PR).

While rebasing I found a defect in this PR worth fixing before merge, so the force-push also changes behavior. source_of() decided built-in vs user-file by testing whether <name>.md existed in any of the three prompt dirs. But context::load() calls prompts::ensure_global(), which seeds every embedded prompt into the global prompts dir on first run, so code.md always exists there and the field reported user_file for every prompt: BuiltIn was unreachable in production. The unit tests missed it because they point ZS_DATA_DIR at an empty temp dir and never call ensure_global().

It now compares content instead: read the highest-precedence on-disk copy (mirroring load()'s override order) and report UserFile only when that text differs from the compiled-in one, or has no compiled-in counterpart. Added three tests, one of which calls ensure_global() first and pins code as BuiltIn (it fails against the old logic).