Autoloop PR/issue should report cumulative performance improvement

## Problem

Autoloop programs that optimize a metric over many iterations don't report the cumulative improvement anywhere visible — not in the PR body, not in the issue, not in the state file summary. A reader of [tsessebe PR #297](https://github.com/githubnext/tsessebe/pull/297) or [issue #189](https://github.com/githubnext/tsessebe/issues/189) can see the *current* best metric but has no idea where the program started or how much progress has been made.

This was observed on the `tsb-perf-evolve` program (41 iterations, fitness ratio optimization). The PR body says "Current best metric: 21.048" but never mentions the starting value or total improvement.

## Root causes

Three gaps in `autoloop.md`:

### 1. No `Initial Metric` in Machine State

The state file schema tracks `Best Metric` but not the baseline metric from the first accepted iteration. Without this, there's no reference point to compute cumulative improvement.

**Fix**: Add an `Initial Metric` field to the Machine State table schema. Set it on the first accepted iteration and never overwrite it.

### 2. PR body template doesn't include cumulative improvement

Step 5c says to update the PR body with "the latest metric and a summary of the most recent accepted iteration" but never specifies showing start-to-finish improvement.

**Fix**: Update the Step 5c PR body template to include: `"Fitness: {best_metric} (started at {initial_metric}, {improvement_pct}% improvement)"` or similar.

### 3. `pending-ci` fitness values never flow back into PR/issue

When the sandbox can't run the evaluation command (e.g., `bun` not available), iterations are pushed as `pending-ci`. After CI runs the benchmark and produces a fitness number, that result never flows back into the PR body, issue status comment, or state file. The PR stays stuck showing the last sandbox-measured metric.

**Fix**: Either (a) add a post-CI callback step that updates the PR body/state file with the CI-measured fitness, or (b) at minimum, have the next iteration read CI results from the previous run and update the state file retroactively before proposing a new change.

## Expected behavior

A reader of any Autoloop PR or program issue should be able to see at a glance:
- Where the metric started (initial/baseline value)
- Where it is now (current best)
- Total improvement (absolute delta and percentage)
- A brief improvement trajectory (e.g., in the PR body or status comment)

## References

- Observed on: [githubnext/tsessebe PR #297](https://github.com/githubnext/tsessebe/pull/297), [issue #189](https://github.com/githubnext/tsessebe/issues/189)
- State file: [`tsb-perf-evolve.md`](https://github.com/githubnext/tsessebe/blob/memory/autoloop/tsb-perf-evolve.md)
- Workflow definition: `.github/workflows/autoloop.md`