pipeline: relax hero floor to 2, keep ~4 as the soft aim

The real failure mode is 0-1 images (wall of text); two heroes read
fine on the 1-2 column layouts. Keep ~4 as the aim — prefer
hero-bearing candidates among equal fits, more images for bigger
bundles — since 2 among 8 items still looks thin on the wide
4-column grid. Restores the 3-item minimum bundle size.

Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL

.claude/skills/build-bundle/SKILL.md

@@ -14,11 +14,11 @@ Invoke from inside the `unique.rzen.dev` repo.
 
 - **Interactive (default when a user is present)**: propose the cluster, wait for sign-off (step 2), then write.
 - **Autonomous (`auto` argument, or when running unattended — e.g. invoked by `/daily-pipeline`, a scheduled run, or any headless session)**: skip the confirmation gate. Apply the same quality bar, write directly, and record the would-have-been proposal in the log so the user can veto after the fact. Prefer the *tighter* theme when torn — an unattended run should err conservative.
-- **Weekly catch (`weekly` argument)**: a date-based safety valve, not a themed bundle. Collect every `shortlist: true` find from the last 7 days that appears in **no** existing bundle's `items[]`; if fewer than 3, widen to all unbundled finds from the window. Title: "The week's catch — <Month D–D>". Blurb names the week, not a theme. Cap at 8 by shortlist rank. The hero floor (see below) applies here too, but widen rather than hold: if the window yields fewer than 4 hero-bearing items, pull in older unbundled hero-bearing finds until the floor is met — the weekly exists so finds don't stay invisible, so holding it defeats the point. Only if the *entire* unbundled pool can't muster 4 heroes, skip and report. If nothing qualifies, do nothing and report that. Purpose: no shortlisted find stays invisible for weeks awaiting a theme (finds are only reachable through bundles).
+- **Weekly catch (`weekly` argument)**: a date-based safety valve, not a themed bundle. Collect every `shortlist: true` find from the last 7 days that appears in **no** existing bundle's `items[]`; if fewer than 3, widen to all unbundled finds from the window. Title: "The week's catch — <Month D–D>". Blurb names the week, not a theme. Cap at 8 by shortlist rank. The hero floor (see below) applies here too, but widen rather than hold: if the window yields fewer than 2 hero-bearing items, pull in older unbundled hero-bearing finds until the floor is met — the weekly exists so finds don't stay invisible, so holding it defeats the point. Only if the *entire* unbundled pool can't muster 2 heroes, skip and report. If nothing qualifies, do nothing and report that. Purpose: no shortlisted find stays invisible for weeks awaiting a theme (finds are only reachable through bundles).
 
 ## Hero floor (applies to every bundle)
 
-A bundle must contain **at least 4 items that have a hero image** (`hero.<ext>` next to the find's `index.mdx` — check with `ls src/content/find/<y>/<m>/<slug>/hero.*`). This is purely aesthetic: the grid looks broken with fewer than four image cards. Hero-less finds may still ride along beyond the four. If a proposed cluster can't meet the floor, **hold the bundle back** — don't pad with weak fits and don't write it anyway; the theme waits until a fourth hero-bearing find emerges. In autonomous mode, log the held-back cluster (theme + items + which lack heroes) so the next run can revisit it.
+A bundle must contain **at least 2 items that have a hero image** (`hero.<ext>` next to the find's `index.mdx` — check with `ls src/content/find/<y>/<m>/<slug>/hero.*`). This is purely aesthetic: with 0–1 images the grid reads as a wall of text. Hero-less finds may still ride along freely — only the floor counts. Beyond the floor, **aim for ~4 heroes**: when choosing among equally good theme fits, prefer hero-bearing ones, and let bigger bundles (7–8 items) carry proportionally more images — 2 heroes in an 8-item grid still looks thin on the wide 4-column layout. If a proposed cluster can't meet even the floor of 2, **hold the bundle back** — don't pad with weak fits and don't write it anyway; the theme waits until enough hero-bearing finds emerge. In autonomous mode, log the held-back cluster (theme + items + which lack heroes) so the next run can revisit it.
 
 ## Inputs
 
@@ -37,7 +37,7 @@ A bundle must contain **at least 4 items that have a hero image** (`hero.<ext>`
   3. Compute the **unbundled pool**: every find that does NOT appear in the `items[]` array of any existing `src/content/bundles/*.mdx` bundle. The seed slugs may or may not be in this pool — they're the seed regardless.
   4. From the unbundled pool, find every entry that fits the inferred theme. Use frontmatter signals (tags, topics, subtitle) and body content. Don't reach: three weakly-fitting items is worse than zero. If a candidate's fit feels strained, drop it.
   5. Combine `seed + qualifying unbundled matches`, deduped. Cap at **8 total** (per sizing notes below). If the combined list exceeds 8, drop the weakest fits — never drop a seed slug; if the seed itself is more than 8, ask the user which to drop.
-  6. Check the hero floor: if fewer than 4 items in the final list have a `hero.*` file, prefer swapping in hero-bearing theme matches; if the floor still can't be met, report that the bundle is held back (interactive: tell the user and let them decide; auto: hold and log).
+  6. Check the hero floor: if fewer than 2 items in the final list have a `hero.*` file, prefer swapping in hero-bearing theme matches; if the floor still can't be met, report that the bundle is held back (interactive: tell the user and let them decide; auto: hold and log).
   7. Produce one cluster (theme phrase + final items list). Continue to step 2 to confirm before writing.
 
 - **If the user provided a theme phrase only:** scan all finds, surface the 8–12 that best fit the theme.
@@ -113,7 +113,7 @@ Example body:
 ### 4. Verify
 
 - Confirm every slug in `items[]` exists at `src/content/find/<year>/<month>/<slug>/index.mdx` (use `find src/content/find -type d -name "<slug>"` to locate). If any is missing, abort and report — never reference a non-existent find.
-- Confirm the hero floor: at least 4 of the `items[]` folders contain a `hero.*` file.
+- Confirm the hero floor: at least 2 of the `items[]` folders contain a `hero.*` file.
 - Run `npm run build` to confirm the new bundle compiles.
 
 ### 5. Update the log
@@ -132,7 +132,7 @@ Items referenced:
 
 ## Notes
 
-- Default to 5 finds per bundle, up to 8 if the theme really supports it. Beyond 8 the grid stops being scannable. The effective minimum is 4 (the hero floor) — a tight 3-item theme waits until a fourth hero-bearing fit emerges.
+- Default to 5 finds per bundle. Use 3 for tighter themes, up to 8 if the theme really supports it. Beyond 8 the grid stops being scannable. Whatever the size, the hero floor (≥2 hero-bearing items, aim for ~4) applies.
 - The body is for the audience, not the user. Never write "pick one", "candidates", "today's pick" — those are internal terms.
 - Never edit existing find files when building a bundle. The find body and metadata are stable; the bundle only references them.
 - If a find that's referenced has been graduated to a review (its frontmatter has `promotedTo`), it can still appear in the bundle — the grid card simply links to the find page, which itself links forward to the review.