[puretensor]

Title: fix: report the latest distill outcome without timestamp ties

Severity: med

Failure scenario: record a successful distill.run at 12:00:00.100000Z, then distill.failed at 12:00:00.200000Z. GET /metrics reports pt_distill_last_run_ok 1 even though the latest run failed. The query truncates both timestamps to whole seconds and treats the tie as success, suppressing the current-failure signal.

Select the latest recorded outcome by the journal's monotonic id. This preserves the no-failures default, reports the later failure, and clears the gauge on a subsequent successful run.

Regression: crates/ptask-server/tests/distill_outcome_order.rs drives the actual HTTP router over a disposable SQLite database, covering the initial state, success followed by failure within one second, and subsequent recovery.

Exact red/green command (same command before and after the fix):

env PATH=/usr/bin:/bin:$HOME/.local/bin:$HOME/.cargo/bin:/home/puretensorai/.local/bin CARGO_HOME="${CARGO_HOME:-$HOME/.cargo}" RUSTUP_HOME="${RUSTUP_HOME:-$HOME/.rustup}" CARGO_TARGET_DIR="$PWD/target" HOME="$(mktemp -d)" cargo test -p ptask-server --locked --test distill_outcome_order

Against unchanged origin/main production code: test result: FAILED. 0 passed; 1 failed (gauge 1 instead of 0). With this fix: test result: ok. 1 passed; 0 failed.

Gate: ran cargo fmt --all -- --check && cargo clippy --workspace --all-targets -- -D warnings && cargo test --workspace --locked. Formatting passes. Clippy has only the two recorded baseline clippy::nonminimal_bool errors in crates/ptask-cli/tests/terminal_safety.rs:78 and crates/ptask-cli/src/ui.rs:946 (exit 101). Existing assertions are unchanged. The separately executed cargo test --workspace --locked passes (exit 0). All runs use an isolated HOME, restricted PATH, and CARGO_TARGET_DIR=$PWD/target; no native-ml feature build.