fix: CLI summary trend deltas use raw delta×100 instead of relative pct_delta

[decko]

Bug

The CLI summary shows absurd trend percentages for non-ratio metrics:

Phase execution time  1382.6s  ▼ -14169%
Tokens / phase        15450.7  ▲ +69766%

The HTML report correctly shows ▲ 11% and ▼ 4% for the same metrics because it uses pct_delta (relative percentage change). The CLI path still multiplies raw delta by 100.

Fix

Update cli_summary.py to use sd.pct_delta instead of sd.delta * 100 — matching the HTML template logic.

[decko]

Spec

Problem

_delta_indicator() in src/raki/report/cli_summary.py (lines 428, 431) computes the displayed percentage by multiplying the raw delta field by 100:

pct = f"+{abs(sparkline.delta or 0) * 100:.0f}%"

SparklineData.delta is the raw signed difference between the latest and prior metric values (e.g. 1382.6 - 15551.6 = -14169.0 for phase execution time in seconds). Multiplying by 100 treats it as a ratio and produces absurd output like ▼ -14169%.

SparklineData.pct_delta already holds the relative percentage change (abs(delta) / abs(prev) * 100) computed correctly in build_sparkline_data(). The HTML template uses pct_delta and correctly shows ▼ 4% for the same metric.

Fix

Replace sparkline.delta * 100 with sparkline.pct_delta in _delta_indicator(). The field is already a percentage — no multiplication needed.

Handle the pct_delta is None case (when the previous value is zero, relative change is undefined) by falling back to no percentage label, matching the HTML template's behavior.

Acceptance Criteria

• _delta_indicator() uses sparkline.pct_delta instead of sparkline.delta * 100
• When pct_delta is None, the arrow renders without a percentage (e.g. ▲ instead of ▲ +0%)
• Capped at >999% to match HTML template behavior for extreme values
• Unit tests cover: normal up/down, flat, pct_delta=None, and pct_delta >= 1000 cases
• No regression in existing CLI summary output tests
• Towncrier fragment 364.fix

---

Plan

Step 1 — Update _delta_indicator() in cli_summary.py

File: src/raki/report/cli_summary.py

Replace the pct computation in _delta_indicator():

Before:

pct = f"+{abs(sparkline.delta or 0) * 100:.0f}%" if sparkline.delta is not None else ""

After:

if sparkline.pct_delta is not None:
    raw = sparkline.pct_delta
    pct = f"+{raw:.0f}%" if raw < 1000 else "+>999%"
else:
    pct = ""

Apply the same pattern for the "down" branch (replacing - ... * 100 with -pct_delta, capped at <-999%).

Also update the docstring: change "Percentage is formatted as an absolute change (× 100)" to "Percentage is the relative change (pct_delta)."

Step 2 — Add/update unit tests

File: tests/test_cli_summary.py (or nearest existing test for _delta_indicator)

Add parametrized cases:

• direction="up", pct_delta=11.0 → ▲ +11%
• direction="down", pct_delta=4.0 → ▼ -4%
• direction="flat" → =
• direction="up", pct_delta=None → ▲ (no percentage)
• direction="up", pct_delta=1500.0 → ▲ +>999%
• direction="down", pct_delta=1500.0 → ▼ -<999%

Confirm no existing test asserts the broken delta * 100 behavior.

Step 3 — Towncrier fragment

File: changes/364.fix

CLI trend delta indicators now show the correct relative percentage change instead of the raw delta multiplied by 100. Metrics like phase execution time and token counts no longer display absurd values such as ``▼ -14169%``.

Estimated scope

• ~15 lines changed in cli_summary.py
• ~20 lines of new tests
• 1 changelog fragment