tachyon MCP server

Four additions for render machines, all read-only like before, plus clearer crash and job reports.

**A Tachyon-only stop is diagnosed from what happened before it**
- When Tachyon stops but the PC keeps running (same boot), only evidence up to the stop counts. Events logged later (an app crashing ten minutes after) go to a separate `after_events` list ("After the stop · not related") and are never evidence. After a restart, application crashes logged after the stop are set apart the same way; the restart's own reports at the next boot still count.
- No RAM, BIOS, PSU or GPU-overclock causes for a Tachyon-only stop without direct evidence (WHEA hardware errors or GPU driver errors just before): the RAM speed, the BIOS start time and the clock belong to the boot, not to this stop.
- Without a crash of Tachyon (its crash log or a Windows Application Error for tachyon.exe) the verdict is "Tachyon was closed at 21:38. Your PC kept running.", a neutral note, not shown on the main view. A real crash of Tachyon reads "Tachyon crashed at ..."; a GPU driver error just before points to the driver.
- Reports saved by an older version are revised on load with these rules (the 26 Oct 21:38 report, which blamed later CC_Engine_x64.exe crashes and EXPO/BIOS from the earlier boot, now reads "closed, nothing to do").
- New `Report::verdict()` (what happened, is the PC OK, what to do) and `severity()` (note / warning / critical).

**Layout: verdict first, details on demand**
- Crash report: one-sentence verdict, then the top cause as a card with plain words (EXPO/XMP explained) and one action; other causes folded; evidence, peaks and timeline behind one row. Times in their own column, text wrapped (no ellipsis), at most three bullets per block, numbers and units aligned in the peaks grid, colour only for meaning (neutral note, amber warning, red critical). Events after the stop shown apart. Sections give way to "in the export" lines when the screen is short (checked at 100, 125 and 150 % scaling).
- New Job report view (menu "Last marked job report", "Job report ›" on the main view and in History): verdict, the numbers in an aligned grid (GPU, power, VRAM, energy, temperature, power-cap and heat-limit time), the PCIe link in one sentence, VRAM by process with bars (three shown, the rest folded), clock-limit reasons and earlier jobs folded.
- Main view: one type scale (12.5 body, 14 values and verdicts, 9.5 labels), wrapped sentences, the limiter shows its first reason "+N" instead of a cut word, a PCIe link row (neutral when full or saving power at idle, amber when slow), process rows drop whole details that do not fit instead of cutting them. The session timer row gave way to the PCIe link.

**Status, comparisons and plain words**
- One status scale for every verdict (`status.rs`): all fine (green dot), worth a look (amber), act now (red), always with its words ("ALL FINE") next to the dot. Crash reports, job reports, the pre-render check, the bottleneck line and the PCIe link use the same mapping.
- Cause shares are percentages only when two or more causes compete (the second has 20 % or more and the top one is under 80 %); otherwise a word: certain, likely or possible. No share bar for a single cause. The voice line follows the same rule.
- The job report compares with the previous run of the same job only (same name, run counters ignored; never a different job): speed, average GPU, VRAM peak, average power and energy as deltas, or "First run, nothing to compare." Also in History and in `get_job_reports` (`vs_previous`, `compared_with`).
- Plain words: "At power limit 55%" (share of the job's time, normal under full load) instead of "power cap 169 s"; "slowed by heat"; clock-limit reasons explained in one sentence each; the PCIe link glossed ("Gen5 x16 (PCIe 5.0, 16 lanes)"). The limiter on the main view says "power limit" in plain text (normal), heat or a power brake in amber.
- The "Tachyon only reads..." note sits behind a small (i) in t