Improve experience when shell output exceeds runtime streaming limits in agent host sessions

Follow-up from https://github.com/microsoft/vscode/issues/328299 — streaming was fixed by #328351 and layout by #328592; this covers what the user is left with when a command's output exceeds what the runtime streams.

**Behavior per output size (current):**

| Output size | While streaming | On finish |
|---|---|---|
| < ~10KiB | full output appends live ✅ | full output ✅ |
| ~10–20KiB | first ~10KiB + one truncation marker, no flicker ✅ (bytes past 10KiB aren't streamed by the runtime) | final preview applied; completeness in this band depends on model-facing caps |
| > ~20KiB (runtime spill) | head + marker + stitched rolling tails accumulate ✅ | ❌ replaced by a 500-char stub — this issue |

**Repro:** in a Copilot agent host session (non-pty built-in shell) with the output preview expanded, run something producing >20KiB, e.g. `gh run watch --repo microsoft/vscode --exit-status` for ~2 minutes, or `for i in $(seq 1 3000); do echo "line $i padding padding padding"; sleep 0.005; done`

**Today's experience at completion:**
1. The card replaces everything that streamed with the runtime's completion text — strictly less content than was on screen a second earlier, and the same after window reload.
2. That text does say the output was too large, but it's written for the model, not the user: tool guidance like "Consider using grep, head/tail, view_range…" and a saved-file path as inert text you'd have to select and copy by hand.

**Why it happens:** `shell_exit` is currently omitted from external completion events, so `outputTruncated` arrives unknowable and `NonPtyShellTerminalStreams.completeToolCall` treats the completion preview as authoritative and rewrites the channel; independently, the retained static result the card renders post-completion and post-reload only ever carries that preview. This is also why model-facing prose appears in the card at all — the structured `shell_exit` carries a clean raw-output preview plus an explicit truncated flag, while the fallback can only harvest the text written for the model.

**Looking ahead:** once the `shell_exit` API is restored to external sessions, the completion preview becomes a clean 500-char head of raw output — meaning the "Output too large…" text disappears from the card entirely, and with it the only hint that anything is missing. At that point a deliberate truncation experience (clear indicator, access to the full saved output) stops being polish and becomes required, or truncation turns fully silent.