fix(ai): analyse the actual session, not stale/benign logs — 0.28.1

The user ran a game ~20s with no crash but the AI dredged up old log lines,
guessed the wrong game, and gave Windows advice. Fixes:
- Prompt now includes the real game name + capture duration + outcome (clean vs
  crash), so the model uses the known game instead of guessing from log paths.
- gamelogs.collect(since=…): scope Steam-console lines by timestamp and skip a
  stale per-app Proton log (mtime before the session) — no unrelated past run.
- ai_knowledge: flag benign Steam/Proton lines (libnvidia-ml.so.1 assertion,
  routine minidumps, "fork without exec") as non-causal.
- System prompt: Linux-only steps (no "run as administrator"); don't manufacture
  a problem on a clean run.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -5,6 +5,18 @@ All notable changes to RigDoctor are recorded here. Format follows
 (`MAJOR.MINOR.PATCH`, pre-1.0). `__version__` and `pyproject.toml` must match the git
 release tag (so the auto-updater, D18, can compare versions).
 
+## [0.28.1] - 2026-05-22
+### Fixed
+- **AI explanations were misreading stale/benign logs.** Three fixes so the model analyses the
+  *actual* session: (1) the prompt now states the **real game name, capture duration, and
+  outcome** (clean vs. crash) so the model stops guessing the game from log paths; (2) game logs
+  are **scoped to the session window** (Steam-console lines filtered by timestamp; a stale
+  per-app Proton log from an earlier game is skipped); (3) the reference KB flags common
+  **benign** Steam/Proton lines (`libnvidia-ml.so.1` assertion, routine minidump uploads, "fork
+  without exec") so they aren't reported as the cause. The system prompt also forbids
+  Windows-only advice (no "run as administrator") and tells the model not to invent a problem
+  when the run was clean.
+
 ## [0.28.0] - 2026-05-22
 ### Added
 - **AI explanations now include recent game logs.** When you press "Explain with AI" on a