feat: per-plugin cost trend + threshold warnings (closes #201)

[vivekchand]

Closes #201

What

Extends the existing per-plugin cost attribution with the two missing sub-features from GH#201:

1. Trend over time — detects which plugins are getting more expensive by comparing the last 7 days vs the prior 7 days of cost share
2. Threshold warnings — alerts when a plugin's cost share exceeds a configurable percentage (default 50%)

How

New: _compute_plugin_trend(plugin, plugin_daily_stats)

Computes trend direction (increasing / decreasing / stable) for a given plugin by comparing its average daily cost-share % in the last 7 days vs the prior 7 days. A >20% relative delta is flagged as increasing or decreasing.

Enhanced: GET /api/usage/by-plugin

Each plugin row now includes a trend field. A warnings array is emitted listing any plugins whose cost share exceeds the threshold (default 50%, configurable via ?threshold=N).

New: GET /api/usage/by-plugin/trend?days=14

Returns a per-day time-series for every plugin in the requested window (1–90 days). Shape:

{
  "days": ["2026-03-20", ...],
  "plugins": {
    "exec": [{"day": "2026-03-20", "tokens": 120, "cost_usd": 0.001, "calls": 3}, ...],
    ...
  }
}

This powers frontend trend charts without any additional backend work.

Data collection

_compute_transcript_analytics now builds a plugin_daily_stats dict (day → plugin → {tokens, cost, calls}) alongside the existing aggregate plugin_stats.

Tests

6 new tests in TestPluginTrend:

• test_by_plugin_returns_200
• test_by_plugin_structure — verifies plugins + warnings keys
• test_by_plugin_row_has_trend — verifies trend is one of valid values
• test_by_plugin_trend_endpoint_returns_200
• test_by_plugin_trend_structure
• test_by_plugin_trend_days_param — verifies ?days=7 returns exactly 7 days