In flight, 2026-09-15

Everything not finished, grouped by who it waits on. Three things decide most of it: one PR about to go out, a web build waiting for Steve's go, and a performance question heading for the Roc Zulip. Below that come what landed today, the known gaps nobody has scheduled, and what is parked.

Waiting on me

Nothing is mid-flight on my side.

Waiting on upstream

PRs 147 to 152 are open with Damian, all read cold before sending. PR 152 went out after two cold reads. Its record is codex-zig-ladder/outbound/SENT-pr152-raytracer-nearest-first.md.

PR	what
147	zig plug: the deck report goes to fd 3, and only when fd 3 is open
148	fat16 test fixtures: the second FAT matches the first
149	Raytracer: shading and the camera in the units of Real geometry
150	GuiOS: a sidebar click selects the button under it, and the next frame draws the new pane
151	Fireworks: rnd wraps, as the hash beside it does
152	Raytracer: the nearest object is found by its distance, and only its hit is built (the zig plug's render 12% faster, same pixels)

u61-candidate (~/showell_repos/cobblestone-u61) is Update 60 plus those five and the Raytracer change. Every roc-apps build uses it, by your rule that we emit assuming our fixes land. When Update 61 arrives, the next candidate is cut from it by the same formula, and whatever Damian took drops out.

Waiting on Steve

1. The web umbrella: designed, decided, not built (you said to wait). - The design is notes/roc-web-umbrella.md. - The site is "Cobblestone Roc Projects": one Caddy, a path per app. - Dev on :9210, open; staging on :9200, written only by publish.sh; prod at https://roc.lynrummy.com, deployed from this box after sign-off. - :9201 and :9203 become redirects. - Step 0 is installing Caddy here.

2. BASIC's first publish. Its public URL is still the preview port (:9203/basic/), which shows whatever basic/build.sh last built. The design calls for a build you check first.

3. The published demos are older than today's emitter.

demo	where	built
Safari	:9201 (public)	published 2026-09-11
the GPU gallery	:9204	published 2026-09-12, from cobblestone-u58
the games	:9205	published 2026-09-12, from cobblestone-u58
Safari's preview	:9203/driving/	built 2026-09-12

• Since then rocemit changed bit-shr (to sar), square roots and forwarder calls, and the default root moved to the candidate.
• Safari's emitted Roc changed text in Blit and Mountains only, and its 54 units pass, but its wasm has not been rebuilt.
• Each republish is a build, an eye test and a publish.sh, when you want one.

4. LLVM for the pages. LLVM is welcome where builds are short; say every build's compile time. - The ray bench builds with LLVM in 2 s and runs ten to sixteen times faster than the dev backend. - Whether the framebuffer page's own builds should use --opt=speed depends on their compile times, which have not been measured.

5. The Zulip question. You want one polished question about the dev backend's stack moves and about seeing where copies go. The notes are notes/zulip-stack-moves-question.md. - The evidence so far: 86% of the hottest function's instructions move a value to or from a stack slot. - Dev is sixteen times LLVM on the closest-hit walk. - A Codex change the zig plug rewards by 12% barely registers under dev. - Still to gather before asking: - a small Roc program, with no Codex or rocemit in it, that shows the same thing; - the same comparison on Renderer3D's hot row loop; - whether wasm under dev shows the ratio native does.

Landed today, for orientation

what	where
codexrun: plain Integer overflow traps, bit-shr is arithmetic, a Real prints as x86 prints it	rust-codex-compiler e43d269
rocemit: bit-shr is Roc's arithmetic shift	a613aff
rocemit: geo-sqrt and quat-real-sqrt are Roc's sqrt; raytrace-on-screen's page frame 96-101 ms to 61-69 ms in Node	ad24db4
rocemit: a call to a forwarder that calls back is written as the call it makes; GlobeDemo runs in the page	fdd45bd
roc-apps builds from u61-candidate; GuiOpening's clicks land; raytrace-on-screen is lit; Fireworks joins the page	roc-apps 9eca5c4
framebuffer/bench/raybench.sh, the native ray bench (OPT=speed for LLVM)	roc-apps 3702572, 57a1ca0
the raytrace case study	notes/raytrace-case-study.md
the unused fireworks-rnd-wrapping branch deleted	NewRepository

Known gaps, not scheduled

• The globe can't be dragged. It reads codex-vm's relative mouse packet at 28684, which the framebuffer platform does not model. The arrows and the keypad work.
• Fireworks' F1 mode (how each frame clears) is wired to the keyboard but has not been exercised.
• Two Cobblestone apps still don't run in the framebuffer page.
• apps/circuits: BitmapFont and the app import each other, and roc build crashes on the cycle.
• apps/c64: it plays its SID through the HDA sound card.
• rocemit's half of "a Text is CCE units" (List(U8)) is not built, so the ladder keeps four CCE failures.
• Does the zig plug write ~ as IEEE ==?
• codexzig's output for geo-sqrt-loop ends on next == guess, where x86's ~ means within four ordinals.
• If that is the rule, a Newton loop that settles between two neighbouring values, or on a NaN, never ends under the zig plug.
• It is a question until ZigEmitter's rule for ~ has been read.
• Two candidate x86 print bugs, found by reading code, need one bare-metal print on cobblestone-qemu before they can be sent.
• Tier-1 codes 128 to 383 print 64 code points high.
• A negative tier-2 delta is added unsigned.
• Checks not rerun.
• The resolver ledger (cobblestone-qemu FINDINGS.md 3c) is untested.
• The parked Rust arm's own gates have not been rerun against Update 60's oracle.
• Roc's debug checker is quadratic in literal leaves per file. A one-line fix is identified, unverified and unreported; that is your call.
• Renderer3D is the scene demos' version of the raytrace study. Nearly half of scene-on-screen's frame is one function, r3d_scan_cols_sh!, and a minimal program for it is the obvious next case.

Parked, by your call

• BASIC in Roc. Next when it resumes: better coverage on the games (55 of 99).
• The Rust compiler's type-engine work.
• The machine essay's rewrite.