feat(marking): borderline detection — flag rating vs derived-band disagreement

[hyperpolymath]

Stacked PR (4 of 4 follow-ups). Base = `feature/composer-grid-component-mapping` (PR #54). Merge order: #51 -> #52 -> #53 -> #54 -> this.

Summary

When the tutor's ordinal rating for a component disagrees significantly with the band derived from the per-component numeric grid, the composer flags it so the tutor can sanity-check before generating feedback.

Severity heuristic

• Ratings ranked 0-5: `missing=0, serious_issue=1, weak=2, adequate=3, sound=4, strong=5`
• Bands ranked 0-5: `fail=0, bare_fail=1, adequate=2, fair=3, good=4, excellent=5`
• `|rating_rank - band_rank| >= 2` → flagged
  • `>= 3` step gap → `:severe`
  • `== 2` step gap → `:mild`
• Sign of gap → `:over_rated` (rating > derived) or `:under_rated`

A 1-step gap (e.g. "strong" rating + good-band numeric) is not flagged — that's normal rating-vs-band variance.

What ships in this PR

• `borderline_check/2` public API on the composer — input rating + aggregate, output flag map or `nil`
• `compose/1` result gains `:borderlines` (`%{component_id => flag}`)
• LiveView
  • Amber banner above the per-component cards when any flags are present
  • Flagged per-component cards get amber styling + one-line explainer (`"rating: Strong, numeric: bare fail (severe over rated)"`)

What this does NOT do

• It does not muddy the `did_well` or `improve` prose itself. The flag is a meta-observation surfaced in the UI; the composed feedback stays clean. (Easy follow-up if you want it inline in prose.)
• Thresholds (2 step / 3 step) are hardcoded. Configurable thresholds per rubric would be a small follow-up.

Tests

• 45 total, 11 new
• severe over_rated (strong + fail-band), severe under_rated (weak + excellent-band; missing + good-band)
• mild over_rated (2-step gap)
• no flag for 1-step gap (strong + good, sound + fair)
• no flag for ranks within tolerance
• `nil` for missing aggregate or unrecognised rating
• `compose/1`: per-component flagged when disagreement; empty map when ratings agree or no numeric data

Test plan

• 45/45 pass via standalone elixirc + ExUnit
• LiveView syntax-checks
• In browser: enter Q1=5/25 with rating "strong" on reflective → component card turns amber, banner shows "1 component shows rating ↔ numeric disagreement"
• Adjusting either the rating to "weak" or the mark upwards clears the flag

🤖 Generated with Claude Code