fix: parse summarized usage reports and AI-credit billed Copilot rows

[austenstone]

Stacked on #104. This is the one that matters.

The task was to check whether the parser still matches what GitHub emits today. It doesn't. Two real drifts, both silent-ish, both confirmed against current GitHub docs and real export samples.

1. Summarized usage reports were rejected outright

GitHub emits the metered usage report in two shapes:

	detailed	summarized
max range	31 days	1 year
username	✅	❌
workflow_path	✅	❌

Detection required workflow_path, so every summarized export threw Unknown report type. For a tool whose entire job is "drop your usage CSV here", and where summarized is the export you reach for when you want more than a month of history, that's a big hole.

date, product, sku, quantity, unit_type, applied_cost_per_quantity,
gross_amount, discount_amount, net_amount, organization, repository,
cost_center_name
→ Error: Unknown report type

Fixed by keying detection off repository instead — no Copilot report has that column, so it's a clean discriminator. The mappers already default missing columns to '', so summarized files need no separate report type.

While in there: HEADER_SIGNATURES was an object being iterated with Object.entries(), i.e. detection order depended on JS key insertion order. Now an explicit ordered array, which also drops a as ReportType cast.

2. AI credits billing rendered zero usage

Copilot moved from per-user PRU quota to token-metered AI Credits on 2026-06-01. That shipped two months ago. It introduced the ai-credits unit type and the copilot_ai_credit / coding_agent_ai_credit / spark_ai_credit SKUs.

PRODUCT_METRIC_OPTIONS.copilot filtered usage on unitType === 'requests'. Under the new billing model that matches nothing, so the "Usage (PRUs)" card silently rendered 0 — no error, no empty state, just a wrong number. Worst kind of bug.

Now filters on "not billed per seat" (!== 'user-months') rather than an allow-list, so the next unit type GitHub invents doesn't break it again. Relabeled to just "Usage" since it's no longer PRUs.

Premium-request detection also moved off the aic_* columns onto model/exceeds_quota/total_monthly_quota — same reason, those are stable across the billing transition.

Tests

7 new regression tests: summarized detection, premium-request without aic_* columns, token-usage still winning over premium-request, org-level seat activity (which genuinely has no Organization column — confirmed against GitHub's own export source), plus end-to-end parses of a summarized report and an ai-credits row.

Checked and found clean

Detailed metered usage, token usage (19 cols), GHAS active committers, dormant users, and enterprise members all match the app exactly. The legacy pre-enhanced-billing Actions CSV was retired in Nov 2025, so no back-compat needed there.

Verified: 254 tests pass, 0 lint errors, 0 type errors, build succeeds.