Aggregate deterministic structural-improvement sub-score

[JSONbored]

Context

Part of #4737 (parent epic — deterministic tier). Depends on the
complexity-delta and duplication-delta sub-issues (sibling sub-issues in this epic); reuses
codecov/patch's existing coverage numbers and (optionally) slop.ts's existing test-evidence signal.

The problem

This epic's deterministic tier is currently four separate signals (complexity delta, duplication delta,
coverage delta, test-evidence delta) with no combined read. src/signals/slop.ts already establishes the
house pattern for exactly this kind of aggregation — a named weighted-findings model plus a band — and
this sub-issue is that same pattern applied to the positive axis.

Fix

Add src/signals/improvement.ts (mirrors slop.ts's shape deliberately, for a maintainer who already
knows how to read one to immediately understand the other):

• StructuralImprovementInput: the REES findings from the complexity-delta and duplication-delta
  analyzers for this PR, plus a coverage-delta figure (reuse Codecov's own patch-coverage numbers where
  available — this repo already computes a real, precise before/after coverage percentage per PR; don't
  recompute it), plus a test-evidence signal (reuse or adapt slop.ts's existing
  missingTestEvidence/test-file detection rather than re-deriving it).
• A weighted-findings formula, buildStructuralImprovementAssessment, output as
  { improvementScore: number; band: ImprovementBand; findings: SignalFinding[] } — deliberately
  parallel to SlopAssessment's shape.
• Bands mirroring slopBandFor's clean/low/elevated/high pattern, renamed for the positive axis (e.g.
  none/minor/moderate/significant — exact naming is the implementer's call, but must be an ordinal
  band, not a raw percentage presented as fact — see the epic's design constraint 1).

Requirements

• This is deterministic-only, no LLM — matches the tier split the epic requires. The LLM-tier
  contribution (sibling sub-issue) is combined at the surfacing layer (sibling sub-issue), not blended
  into this score.
• Every finding needs a plain-language reason string (mirrors SignalFinding shape already used by
  slop.ts and buildPublicReadinessScore) so the eventual panel row can explain itself without needing
  new sanitizer-unsafe vocabulary.
• Zero applicable findings (e.g. a docs-only PR with nothing for complexity/duplication analyzers to
  measure) must degrade to an explicit "not enough signal" band, not a fabricated none/neutral score —
  don't let an analyzer's silence get misread as "no improvement."

Acceptance criteria

• buildStructuralImprovementAssessment combines all four deterministic inputs into one banded
  result with structured findings.
• Explicit "insufficient signal" handling for PRs where none of the four inputs produced anything
  (tested).
• Full branch coverage, invariant tests mirroring slop.ts's existing invariant-test style (correct
  output per band, no cross-band leakage).
• This module and its output are consumed only by the panel-surfacing sub-issue at this stage — not
  wired into any gate/blocker path (epic design constraint 2).

[linear-code]

JSO-3425