…droom overflow projection

v0.2 shipped per-build sizes shards; v0.3 added the Kconfig configurator;
v0.4 finally reads across the retention window. The Drift tab catches
"package X grew between two specific nightlies" — useful but you need
to already suspect a regression to pick the right two builds. TrendsView
watches the slope: if a package has been growing 4 KB/week for the last
14 days, it surfaces at the top of a leaderboard before the rootfs cap
overflows.

This is the direct answer to PR OpenIPC/firmware#2163's pain point —
"~12 KB of accumulated drift since 2026-05-17 has now tipped it over."
With historical trends, the bisect step that PR had to do becomes
visible from the dashboard.

Data shape

  scripts/prebuild.mts now walks every (build × platform) sizes shard
  once per source after the per-tag download loop, accumulates
  per-platform series into one PlatformAccumulator, and writes a
  per-platform aggregated file at:

      public/data/<source>/trends/trends.<platform>.json

  Schema (matches src/lib/timeseries.ts TrendsFile verbatim):

      {
        schema: 1,
        source, platform, generated_at,
        packages: { <name>: [{build_id, built_at, bytes}, ...] },
        modules:  { <name>: [{build_id, built_at, bytes}, ...] },
        headroom_rootfs: [{build_id, built_at, used_kb, cap_kb, headroom_kb}, ...],
        headroom_kernel: [{...}, ...]
      }

  Every series ships sorted by built_at ascending. Defensive against
  partial shards: missing `headroom` blocks are skipped per build, not
  fatal. New FsHooks.read so the test memFs can serve readback inside
  the same in-memory state the gh-download mock writes into.

Per-platform file: ~16 KB raw / ~1.5 KB gzipped on a board with 31
packages × 2 builds of data. Storage scales linearly with retention;
~80 KB raw at full 90-build retention. Total dist/data/firmware/trends/
across all 96 platforms today: 1.8 MB raw. Well under the plan's
~4 MB-gzipped envelope.

UI

  TrendsView (new "Trends" tab between "Drift vs another build" and
  "Configure (what-if)"):

    * Header summary: builds count + date range covered
    * Two HeadroomChart sections — rootfs and kernel — that draw the
      used-bytes curve against the cap line. When projectOverflow(...)
      returns a non-null projection within 60 days, an amber badge
      surfaces "projected overflow in 7d (2026-06-12)". Linear
      regression: minimise (kb - (slope*day + intercept))^2, solve for
      kb=0. Sub-zero slope only — flat/growing returns null (no
      overflow projected).
    * Controls strip: window selector (7/14/30/90d), packages|modules
      toggle, "min weekly delta KB" filter to surface only the
      noteworthy growers.
    * Growers leaderboard sorted by absolute byte delta in window,
      newest 20. Each row gets an inline Sparkline (180×28 SVG
      polyline, colour by direction: amber=grow, green=shrink), plus
      a click-to-expand large sparkline (760×140) with axis labels.

  Sparklines are hand-rolled SVG — no chart lib. d3-hierarchy is
  already in the bundle (treemap) but doesn't transitively ship
  d3-scale/d3-shape; rather than add ~20 KB of charting deps for
  polylines and axes, the component pads + scales manually. Drop-in
  swap to a charting lib is easy if interactive zoom becomes a thing
  later.

Tests

  tests/trends.test.ts (12 cases):
    * growthInWindow: window filtering, sort-on-input-order tolerance,
      first/last point semantics, < 2 points → null
    * bytesPerDay: window-endpoint slope, insufficient data → null
    * topGrowers: absolute-delta ordering, limit, skip insufficient,
      per-day rate
    * projectOverflow: flat/growing returns null, shrinking-at-known
      rate gives the expected days-to-zero

  tests/bundle.test.ts gains the same-origin invariant for
  trends/...json URLs — the regression guard for the original v0.1
  CORS bug class now covers trends URLs too.

  tests/prebuild.test.ts memFs updated with `read` so the runPrebuild
  cases still pass; defensive `headroom?.rootfs` guards in emitTrends
  let the minimal-shard fixtures run through trends emission without
  exploding.

Numbers

  - 63 active tests (was 46 in v0.3, +12 trends + adjustments)
  - JS bundle: 181 KB raw / 58 KB gzipped (+9 KB vs v0.3; under 250 KB
    perf budget)
  - Initial launch byte budget unchanged: ~60 KB gzipped (trends are
    lazy-loaded on Trends-tab click, ~1.5 KB gzipped per platform)
  - 96 trends files emitted today; 1.8 MB raw total

Roadmap from README

  v0.4 (this) shipped → v0.5 (build-request flow, optional, depends on
  v0.3 fragment) remains. Maintenance backlog (cache eviction, 404 UX,
  drift URL state, picker search, data-window header indicator)
  unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>