Studio: include the backend session log in the support diagnostics report

[LeoBorcherding]

Context: the Windows ROCm / Strix Halo crash reported on Reddit (same host as #8853). This does not fix that crash — it makes the next report of it diagnosable.

Problem

The Python backend aims faulthandler at ~/.unsloth/studio/logs/server/ (studio/backend/run.py:1850), so when the GPU runtime aborts, the Python stack naming the call that died is written there and in no other file Studio keeps.

The support report's == Log tails == section collected tauri.log, tauri.log.1 and the phase logs, and never that directory. A hard native crash therefore arrives as exit code: 3 with nothing about what the process was doing, while the one file that answers it sits on the user's disk unasked-for. That is exactly how the linked report played out: seven minutes of healthy polling, the model dropdown's request fan-out all returning 200, then Backend stdout stream ended with status: exit code: 3 with no traceback and no [stderr] line.

Change

Tail the two newest server-*.log alongside the existing sections, labelled source=backend-session-log. Two files, so a crash and the restart that followed it both land in one report. An absent directory reports backend_session_logs=unavailable rather than erroring — a --no-torch host, or one with UNSLOTH_STUDIO_NO_FILE_LOG=1, simply has none.

They get a smaller budget than the other tails (64 KiB / 400 lines vs 200 KiB / 1000). enforce_report_limit truncates the end of the body, so bytes spent in a section inserted here come straight out of the phase logs below it. A faulthandler dump is a few KB at the end of the file, which is the half read_tail returns, so the smaller cap loses nothing that matters.

append_tail_section keeps its signature and delegates to a _capped variant; no existing call site changes behaviour.

Testing

Four new tests in diagnostics::report::tests:

• newest-first ordering and the two-file cap
• only server-*.log is collected (tauri.log, server-notes.txt ignored)
• absent directory yields empty, no panic
• the section stays inside its smaller budget while retaining the trailing faulthandler stack

Both behaviours mutation-tested: flipping the sort to ascending, and raising SERVER_LOG_TAIL_MAX_BYTES to TAIL_MAX_BYTES, each fail the suite.

cargo test --bins: 367 passed. cargo fmt --check is clean on the changed file (the tree has pre-existing diffs elsewhere, left alone).

Privacy: the new tails go through the same redact_text pass as every other section; report_redacts_footer_collection_warning_paths covers home-path masking.