The target, in the owner's words

"mark as read-only where possible. and strip code and associated data where possible." — 2026-09-17

"so, we're up for 2 umbrella tickets. one for PC platforms - and make sure we can emit minimal code. and second for ESP, same goal." — 2026-09-18

This is an umbrella. Do not claim it — claim a rung.

prio: 70 is a placeholder set by an agent. Umbrella prio is the one number a human sets; this one has not been set by him yet.

What is already true, measured 2026-09-18

Every row verified to actually print hello. real is the code segment with trailing zero padding stripped, because code= is the page ceiling.

The 195 bytes, disassembled in full: 110 B startup prologue (getrlimit on RLIMIT_STACK, clamp to 64 MiB, gettid, build a four-qword thread block, arch_prctl(ARCH_SET_FS) for TLS, record rsp) + 85 B of program (write(1,"hello",5), write(1,"\n",1), exit_group).

The owner's recollection was "like 267 bytes of executable memory" for a Pascal hello world without ansistrings. It is 195. That property has not regressed; it has improved. What regressed is the DEFAULT, and the mechanism is known: PasApplyDefaults defines PXX_MANAGED_STRING unconditionally, so every Pascal program pulls builtinheap and {$H-} cannot reach it.

The rungs

1. Make size measurable. [[bug-t-code-is-page-quantised-so-there-is-no-instrument-for-size-work]]. Nothing below can be graded without it — a 405-byte win reads as zero.
2. Stop paying for facilities the program opted out of. [[bug-a-the-signal-alt-stack-is-32768-bytes-of-unconditional-bss]] is 78% of the bss floor on its own. The readln line buffer — the other 10% — is DONE 2026-09-18, and it was worth DOUBLE what this line said: 8,168 B, not 4,096, because the same buffer was reserved TWICE in an x86-64 image, once by the Pascal driver (BSS_LINE_BUF) and once by the builtin unit (PXXLineBuf), and only the first was ever read. Empty program, x86-64: bss 46,596 -> 38,428. i386 / arm32 / riscv32: 42,316 -> 34,140. aarch64: 42,356 -> 34,188. It is a pointer to a demand-allocated growable block now, so a program that never touches stdin reserves nothing at all. NEXT IN THIS RUNG AND IT IS THE LARGEST ONE LEFT: [[feature-a-the-threadvar-area-is-3072-bytes-of-bss-in-every-program-that-has-no-threadvar]] — 3,072 B, 8% of the hosted floor, held whether or not the program declares a threadvar. Measured x86-64-ONLY (0 on i386, aarch64 and bare ESP), so it is deliberately NOT wired under the ESP umbrella, which is the opposite of the two items above it and is why the ticket leads with the profile table.
3. Make the default converge on the floor. [[bug-a-a-pascal-hello-world-is-63kb-after-emission-size-dce]] — the unconditional PXX_MANAGED_STRING. This is the single largest default-path item and it is worth ~66 KB of code and ~4 KB of data.
4. Read-only constants. [[feature-a-there-is-no-read-only-load-segment-so-nothing-can-be-flash-resident]], which is blocked in turn by the two tickets that make constants actually constant. This is the owner's first directive and on hosted targets it buys page permissions and sharing rather than bytes — the bytes are on the ESP side, which is why the two umbrellas are separate.
5. Strip data, not just code. [[feature-a-unreferenced-class-rtti-keeps-every-method-alive]], [[feature-opt-rtti-emit-on-use]], [[feature-a-an-extern-only-variable-still-reserves-its-storage]]. --dce drops procedure bodies and provably removes zero bytes of data or bss; everything in this rung is the other half.
6. Close the frontend/backend visibility hole. [[bug-a-a-frontend-cannot-see-that-a-backend-calls-library-routines-it-never-mentions]]. A pass that cannot see what the backend emits cannot strip it, and the signal runtime is the worked example: it is backend-emitted machine code, not Pascal procedures, so procs=0 and DCE is structurally blind to it.

The trap this umbrella exists to prevent

code= and SRAM are different quantities and the seat that opened this confused them. On a hosted target code is demand-paged from the file and never all resident; on ESP the bare profile loads code into IRAM. A ticket that reports a code-size win as a memory win is wrong on both targets for different reasons. Say which segment, and say which profile.