Per-bullet feedback — redesign from flat dump to actionable model

[s-annam]

Per-bullet feedback — redesign from a flat dump to an actionable model

Part of M5 · UI Improvements.

Problem

PerBulletFeedback in src/components/Result.tsx renders one BulletRow per detected bullet, each with three check pills (action verb / 8–30-word length / metric). On a real resume this means ~35 rows dumped in a single scroll. It's accurate but unusable: no hierarchy, no prioritization, and it buries the signal ("most of your bullets lack a metric") under 35 individually-rendered lines.

Goal

Replace the flat list with a UI that:

1. Leads with the high-level shape — e.g. "12 of 35 bullets need attention" plus a per-check rollup ("28 missing a metric · 9 too long · 4 weak verb"). The summary line already exists in PerBulletFeedback; promote it to the primary content.
2. Categorizes bullets by failure mode (missing metric / length / weak verb), so the user works one category at a time instead of scanning 35 mixed rows.
3. Encourages drill-down without forcing it — collapse the per-bullet detail behind an expand affordance per category (or a "show flagged bullets" toggle); passing bullets shouldn't render at full weight at all.
4. Gives the user a model to act on — order by impact, show the worst offenders first, make it obvious what to fix next.

Stretch (flag as a separate phase — the high-value play)

Auto-rewrite flagged bullets to address their issues (add a metric placeholder, tighten length, lead with a strong action verb). This is where the real value is — turning a diagnostic into a fix. Likely depends on the in-browser AI rewrite work (M4 · #3-area) for the generative side; the deterministic side can pre-fill structure (verb list, length trim) without a model. Do not bundle the rewrite into the first pass — ship the categorized/summarized read-only redesign first, then layer rewrite on top.

Acceptance (first pass — read-only redesign)

• High-level summary (counts + per-check rollup) is the primary content, above any per-bullet detail
• Bullets grouped by failure category; passing bullets not rendered at full weight
• Per-bullet detail is drill-down (collapsed by default), not a flat 35-row dump
• Worst-first / impact ordering so the user knows what to fix next
• npm run typecheck + npm run build + npm run test green

Out of scope (this issue)

• The auto-rewrite stretch — tracked as a follow-up once the read-only redesign lands.

Files

• src/components/Result.tsx — PerBulletFeedback / BulletRow (likely extract into a features/ component; Result.tsx is already over the 200-LOC guideline)
• src/lib/score/score.ts — BulletObservation shape (read-only; check whether categorization needs any new derived field)

[s-annam]

Implementation Complete — 299978f (PR #49)

Acceptance Criteria

• Per-bullet detail is drill-down, not a flat dump — flagged bullets grouped by failure category, each category in a native <details>/<summary> collapsed by default; passing bullets counted in the summary, not rendered at full weight.
• Worst-first / impact ordering — categories sorted by failing-bullet count descending; within a category, bullets sorted by most checks failed.
• npm run typecheck + npm run build + npm run test green — typecheck clean, build clean, 194 tests pass.
• Auto-rewrite stretch out of scope — not implemented; deferred as a follow-up.

Additional Changes

• Component extraction: PerBulletFeedback + BulletRow + needsAttention moved out of Result.tsx (167 inline lines) into new src/components/features/PerBulletFeedback.tsx (267 LOC). Result.tsx drops ~480 → ~320 LOC, easing the 200-LOC-per-component guideline.
• Summary rollup leads — total bullets + per-check pass/fail counts (missing metric / length / weak verb) surfaced as primary content above the drill-downs.
• needsAttention broadened to OR-of-failed-checks (any failing check flags) for an accurate summary count.
• score.ts untouched — BulletObservation already carried every needed field; no new derived field.

Follow-up: file an issue for the deferred auto-rewrite stretch.

[s-annam]

Implementation Complete — a044a4c (PR #49)

PerBulletFeedback extracted out of Result.tsx into src/components/features/PerBulletFeedback.tsx. The first-pass acceptance was redirected mid-review by design feedback — noting honestly where the shipped result matches the issue and where it deliberately diverges.

Acceptance Criteria

• High-level summary is primary content — rollup ("N of M bullets need attention" + per-check counts) sits above all per-bullet detail.
• Passing bullets not rendered at full weight — passing bullets are counted in a single footer line, never listed as rows.
• [~] Bullets grouped by failure category — intentionally superseded. Per-failure-mode grouping was built (PR feat: extract PerBulletFeedback + drill-down disclosure model #49 first cut) but flattened to one list on review: category grouping duplicated any bullet failing two checks, and the dup also duplicated the rewrite affordance. Failure modes now surface as inline chips per bullet + the rollup counts.
• [~] Per-bullet detail is drill-down (collapsed by default) — changed. Only failing bullets render (each once), so there is no 35-row dump to collapse; the flat list is the detail. Passing bullets stay collapsed (counted only).
• [~] Worst-first / impact ordering — intentionally superseded. Shipped in resume order at the user's direction (worst-first reordering made the list hard to map back to the resume).
• typecheck + build + test green — 243 tests pass; typecheck + build clean; @mlc-ai/web-llm still code-split into its own ~6 MB chunk.

Additional Changes (beyond first-pass scope)

• Auto-rewrite stretch landed (the issue deferred it): the WebLLM "Rewrite" affordance from feat(webllm): per-bullet "Suggest a rewrite" pilot — WebGPU-only, Qwen2-1.5B #39 is integrated — a low-weight inline Rewrite text-link (SVG sparkle, 44 px tap target, aria-label), on failing bullets only, lazy-loading Qwen2-1.5B on click.
• Resolved a semantic rebase conflict — PR feat: extract PerBulletFeedback + drill-down disclosure model #49 predated feat(webllm): per-bullet "Suggest a rewrite" pilot — WebGPU-only, Qwen2-1.5B #39 landing the RewriteButton in this same component; the extraction was reconciled to preserve the rewrite feature rather than silently drop it.
• Follow-up filed (Rewrite UX: one call for all bullets vs per-bullet trigger #50) — single "Rewrite all flagged" call vs per-bullet trigger is left as an open product decision.