[Ticket]: Prompt Budget: coverage-loss panel shows No data instead of 0% when no run lost coverage

## Problem

On the Prompt Budget dashboard (added in #599), the **"Coverage loss — runs with omitted files"** panel displays **`No data`** when no run has lost coverage — i.e. in the healthy case.

Observed live on 2026-08-08 against production: 13 review runs on the `fast` preset, `Files omitted, total = 0`, and the coverage-loss panel rendering `No data`.

That is the *good* outcome being displayed in the same way a broken panel looks. On a dashboard whose entire purpose is a one-time sizing decision, "healthy" and "this query is broken" must not be visually identical — especially in a repo that has been repeatedly bitten by telemetry that silently returns nothing.

## Cause

The panel is a ratio:

```
100 * (sum by (fields_preset) (count_over_time( 0>)))
/ (sum by (fields_preset) (count_over_time()))
```

When no run has `diff_files_omitted > 0`, the numerator produces **no series at all** for that preset. A binary operation with a missing left-hand series yields no result — standard PromQL/LogQL vector-matching semantics, not a bug in itself. The denominator is fine; there is simply nothing to divide.

## Proposed fix

Substitute a zero-valued series carrying the denominator's labels when the numerator is absent — the usual PromQL idiom:

```
100 * (( or ( * 0)) / )
```

**Do not** solve this with Grafana's "No value → 0" panel option. That would be wrong: it cannot distinguish *"runs happened and none lost coverage"* (genuinely 0%) from *"no runs at all in this range"* (genuinely no data). Those are different states and an operator sizing a cap needs to tell them apart.

The query-level fix preserves that distinction correctly:

| state | numerator | denominator | result |
|---|---|---|---|
| runs happened, none lost coverage | absent | present | **0%** ✅ |
| runs happened, some lost coverage | present | present | real ratio ✅ |
| no runs at all in range | absent | absent | **No data** ✅ |

## Acceptance criteria

- [ ] Panel reads `0%` when diff-bearing runs exist and none have omitted files
- [ ] Panel still reads `No data` when there are no diff-bearing runs at all in the time range
- [ ] Per-preset breakdown is preserved (the fix must not collapse to a single unlabelled series — `or on() vector(0)` would do exactly that and is the wrong tool here)
- [ ] Change made in `tools/dashboard-gen/lci_dashboards/prompt_budget.py` and the JSON regenerated, never hand-edited

## Verification required — do not merge on a syntax read

LogQL's support for set operators between metric queries needs confirming for the exact expression used, and this has burned us twice already on this dashboard: #599 shipped four panels whose `count_over_time(...) by (...)` / `sum_over_time(...) by (...)` were outright parse errors, and an `avg_over_time` that silently returned 26666.67 instead of 40000. Both were found only by running the queries.

So: stand up a local Loki (`docker run grafana/loki`), push genuine `tracing`-emitted lines — real output from the actual code path, not hand-written JSON, since hand-writing the absent-field case inverts the result — and confirm all three rows of the table above with real query output. Take the query from the generated JSON, not the Python source.

`tools/dashboard-gen/lci_dashboards/prompt_budget.py`'s module docstring documents the method and the Loki versions already tested (3.1.1 and 3.5.7).

## Context

Follows #599. Not urgent — the panel is misleading, not wrong, and `Files omitted, total` already gives the same signal unambiguously. But it should be fixed before anyone uses this dashboard to actually pick a new `max_diff_chars`.