Anchor Website 15

Fable 5

Summary

THE CUTOVER SESSION: finished the Box2D letter-drop (full rotated video controls + the reset escalation ladder), posted the announcement, and flipped a327ex.com to the engine renderer — the site now lives on Anchor 2 with the Lua server underneath as the SEO layer. Then a long post-cutover hardening tail: content-hashed bundle caching, the video seek/CORS architecture, and several web-only fixes.

Rotated video controls in the drop (the "too hard" feature, done in ~120 lines):

• s14 had rejected in-place controls on fallen video cards as too hard ("controls are axis-aligned so they don't fit a rotated card"). Fresh analysis found three structural facts that made it nearly free: (1) video_controls_geom(r) derives every control rect purely from the card rect; (2) the drop already re-draws whole cards through a rotating layer_push transform, so rotated RENDERING of controls/title/popups costs nothing; (3) the stashed hit-rects (el._rect, _time_rect, _title_rect, _yt_btn_rect) are doc-space during the drop — self-consistent with a doc-space mouse.
• The entire feature reduced to inverse-transforming the mouse into the fallen body's frame: drop_local_mouse = p = center + R(-a)·(m − draw_pos), verified against the engine's [c -s; s c] matrix in layer_push.
• video_controls_update(el, r, mx, my) gained optional mouse params; new drop_video_update() runs the full unmodified controls bar for the active card each fall-frame (seek/volume/speed/CC/time-toggle/title all work rotated) + topmost-wins click routing on other video cards (activate / watch-on-YouTube / placeholder cards). New doc_mouse_position() override in elements.lua keeps draw-side hover visuals correct (set per-body in draw_drop); draw-side clicks are suppressed during the drop (ui_mouse_pressed cleared around the element redraws) — this also fixed a latent ghost-click bug (clicking a video's pre-drop position started playback).
• Fullscreen escapes rotation: the normal axis-aligned player draws over the scene (update + draw gates handle fullscreen during the drop), Esc back to the pile. Owner: "works wonderfully."

The reset mechanic (increment 3) + the announcement test post:

• The [this](#drop) link stays STATIC: draw_text_element records _link_rects with url #drop into drop_links during capture; classify_glyph routes their glyphs to drop.statics, drawn upright + underlined ON TOP of the pile (a settling card can't hide the escape hatch).
• Records refactor: every drop record carries hx/hy (home center) + dx/dy/da (current draw transform — physics-synced in 'fall' via drop_sync_physics, interpolated in 'return' via drop_return_step); draw_drop reads ONLY records. drop_return_begin samples start transforms then DESTROYS all bodies + walls immediately (the tween is pure draw-side); 0.8s smoothstep home, angles unwind ≤ half-turn (Box2D v3 b2Rot_GetAngle is normalized). At p=1 the untouched page resumes; a playing video plays straight through.
• Tried easeOutBack overshoot on request, owner reverted to smoothstep at 0.8s (from 0.6).
• Local test post: found the owner's verbatim announcement text in the s14 vault log (private-session-18.md); realized "the video below (the Box3D one)" = the REAL Box3D announcement message (2026-07-01-114939, ::youtube jr_Fzl2XwKU — Erin Catto's "Introducing Box3D") that sits below the announcement in the feed. Wrote messages/drop-test.md (untracked; convert reads messages/*.md), pulled jr_Fzl2XwKU locally (220MB, the s14 "pre-existing drift" id).

Feel iteration on the drop physics (owner-directed):

• drop_push: random sideways velocity on non-letter bodies, scaled INVERSELY with sqrt(area) (~mass) — heavy cards drift, light avatars fling; 0.4–1.0 magnitude roll, 400 px/s cap, DROP_PUSH=60000. Later boosted 2.5× for videos/tweets and added to divider bars (feather-light → they max the cap).
• The #drop link became SOLID as well as static (one static wall box per fragment — letters pile on the word).
• The Box3D video (DROP_HERO='jr_Fzl2XwKU') launches UPWARD 900–1200 px/s with a guaranteed 4–7 rad/s spin. Tried making the launch radial from the blast; owner reverted to straight-up.
• BLAST EPICENTER: the phrase "absolutely blown the fuck out!" (drop_blast_phrase global) — draw_text_element matches it as a consecutive word-token run (survives line wrap) and records per-word rects; its glyphs explode radially 1400–2000 px/s + a 250 px/s up-kick + drop_spin(30); video spin variation doubled (drop_spin(1.6)).

The escalation ladder (hunt mode + the bomb):

• drop.loops counts completed resets. Loops 10–19 = HUNT: the word falls too — ONE rigid body per link fragment (drop.words; glyphs as fixed offsets drawn rigidly with the block, underline rides it; first version fell as individual letters, owner asked for a solid word block). Reset = click the fallen block (rotated hit-test, HUNT_PAD=10).
• Loops ≥20 = BOMB: clicking the hunted word calls drop_explode instead of resetting — radial velocity kick to every body from the word's position, linear falloff (BLAST_R=700, BLAST_V=2600), the word itself launches in a random direction. Repeatable; reload is the only way back.
• Sidebar caret fix: the disclosure chevrons are layer_line strokes — invisible to the glyph capture — so they vanished during the drop. sidebar_caret became a global stroke-draw; carets are captured into drop_carets and become tiny letter bodies that tumble/tween like everything else.

⭐ ENGINE BUG: the 20-click freeze = shape-pool corruption (Anchor2 859e43d):

• Owner: after exactly 20 clicks (10 loops) everything stopped being interactable. Diagnosis: the per-shape user-data pool in anchor.c was a BUMP allocator (shape_user_data_pool[shape_user_data_count++], MAX_SHAPE_USER_DATA=4096) with NO recycling and (except chains) NO bounds check — b2DestroyBody freed the Box2D shapes but not the slots. ~400 shapes per drop cycle × 10 loops crossed 4096 and wrote past the array into adjacent statics (memory corruption → everything dead, page looked normal).
• Fix: a free list — shape_user_data_alloc() (recycled slots first, then bump, NULL on true exhaustion → clean luaL_error) + shape_user_data_recycle() called from physics_destroy_body (walks the body's shapes, 32-cap convention) and physics_shape_destroy; both world-reset sites clear the free count. REUSABLE for any Anchor2 game that churns bodies.

Staging deploy + ship prep + THE FLIP:

• Touch-scroll modal fix (found pre-deploy): the drain sat in the normal update path; the drop/lightbox/fullscreen branches return before it, so mobile swipes during any modal accumulated in the engine counter and landed as one violent scroll on close. Drain moved to the top of update(), modal branches discard.
• anchor.c committed to Anchor2: 1903561 (glyph capture + layer_draw_glyph_c), then 859e43d (pool fix), later 565ba1b (media base). Renderer committed to a327ex-site (d7cd5e5 + follow-ups). deploy.sh --content --engine shipped everything to staging; owner verified on desktop + phone ("not chugging").
• Ship prep: F9 trigger removed (bind + both input_pressed checks), drop-test.md deleted, the REAL announcement posted via the /msg flow (messages/2026-07-01-213155.md, owner's final wording — he revised it twice in-session; single paragraph, [this](#drop)), renderer synced.
• THE FLIP: sudo cp deploy/nginx-a327ex-engine.conf /etc/nginx/sites-available/a327ex && nginx -t && reload on the VPS. Verified: engine-boot.js injected in root HTML, renderer.data/anchor.wasm 200 at the apex, per-page server-rendered <title>s intact (SEO layer), lazy /data + /media-cache→R2 proxy OK. REVERT = re-apply deploy/nginx-a327ex.conf.

Post-cutover hardening:

• Tab title: the engine's SDL window title ("a327ex site renderer") stomped document.title on web. Boot title → 'a327ex.com'; engine_set_title (existing binding) on home/article/settings navigation matching the Lua site's exact format ("home — a327ex.com", "