Lean library-OS dissolve: gale components must be no_std/no-grow so meld can shared-fuse (single-address-space MCU lowering)

[avrabe]

Correction of the dissolve pipeline (supersedes the synth mis-route)

The canonical dissolve is meld fuse → loom → synth, not wac compose + wasm-tools unbundle (which I wrongly used in PR#88 — it preserves per-component adapters and yields two ~24 KB adapter-laden cores). meld is the fusion stage: it resolves imports, merges index spaces, and generates the canonical-ABI adapter at build time into one monolithic core, which is what loom/synth want.

The actual blocker (= the no_std/no_alloc point)

meld fuse --memory shared --address-rebase (the MCU single-address-space mode) fails:

unsupported component feature: memory.grow not supported with address rebasing

Both gale-app-demo and gale-kiln carry memory.grow (1 each) — the default Rust/wit-bindgen wasip2 allocator. The multi-memory fallback isn't MCU-lowerable (2 memories, 1088 KB each; synth loud-skips the cross-memory copies per #369 — correct behavior).

Fix (gale side)

Build the gale application + host components #![no_std] with a static/bump allocator that never grows (no memory.grow) — the same no_std/no_alloc discipline the payload already follows. Then:
meld fuse --memory shared --address-rebase → single lean core (no per-component adapter, scalar ABI ⇒ no cabi_realloc needed) → loom DCE → synth → one MCU library-OS .o (target the wasm-dist 544 B-class, not 24 KB).

Kill-criterion

"With no-grow components, meld fuse --memory shared still fails / the synth'd single core still exceeds a few KB of adapter" → not fixed.

Supersedes the FIND-BYOOS-006 framing; corrects PR#88. Relates: gale#74, the no_std/no_alloc seam (wit/README), meld fusion.

[avrabe]

Correction (thanks to the steer): the memory.grow is not gale code and the primary fix is meld, not "make our components no-grow."

Located it: cabi_realloc (exported, wit-bindgen canonical ABI) → __rust_alloc → dlmalloc → sbrk → memory.grow. After meld fuse the app↔kernel boundary is internal (0 remaining imports), yet the fused core still exports cabi_realloc ×2 and keeps memory.grow ×2 — meld leaves the vestigial canonical-ABI adapter in place, and being exported it can't be DCE'd. That dead allocator is what blocks --memory shared.

→ Real fix filed on meld: drop the now-internal cabi_realloc/adapter on fusion (scalar external surface) so the allocator+grow DCE. no_std/no-grow components remain a secondary belt-and-suspenders but are not the root cause. Keeping this open as the gale-side tracker for the lean-image milestone, gated on the meld fix.

[avrabe]

Cross-link: the meld-side of this dissolve blocker is now tracked at pulseengine/meld#299 — meld fuse --memory shared --address-rebase correctly fail-louds on memory.grow (SR-37 frozen-layout soundness), so meld needs either a sound bounded-static-heap lowering of memory.grow (general fix) or an actionable diagnostic pointing to the no-grow contract this issue (#89) pursues on the gale side. The two issues are the gale-side (no-grow components) and meld-side (sound grow path) halves of the same single-address-space MCU lowering.

[avrabe]

Upstream root-cause forwarded to pulseengine/wit-bindgen#6 (emit no-grow/elided cabi_realloc for scalar-only worlds) + meld#298 (drop the vestigial adapter on fusion). gale#89 stays the consumer tracker; clears when either lands and meld fuse --memory shared yields a grow-free lean MCU core.

[avrabe]

Validated against pulseengine/wit-bindgen@integration/embedded-rt-no-grow — the lean MCU dissolve is unblocked.

Built gale-kiln for wasm32-unknown-unknown with cabi-realloc-extern: memory.grow = 0, dlmalloc/sbrk = 0, cabi_realloc routes to an imported __cabi_arena_realloc. Dissolved (loom → synth cortex-m4f): 972 B .text (11 fns) vs ~24,848 B for the wasip2 adapter-laden version — ~25×.

Working path for the library-OS image: grow-free core (wasm32-unknown-unknown + cabi-realloc-extern) → loom → synth → 972 B .o, linked against a TCB providing __cabi_arena_realloc (fixed arena, trap-on-exhaustion). The gust/BYO-OS TCB owns that arena (one policy per image).

Remaining gale-side work to close this: (1) the dissolve build switches the components to wasm32-unknown-unknown + cabi-realloc-extern (the wasip2 build stays for the wasmtime/compose demo); (2) the TCB provides __cabi_arena_realloc. Caveats forwarded to wit-bindgen#4: the feature is gated out for target_env=p2, and wasm-tools component new rejects the env arena import (so meld-shared-fuse on grow-free inputs is separately blocked; not needed for the core-dissolve path).

[avrabe]

Adopted + validated — lean grow-free dissolve works (PR #90).

Pinned both component crates to wit-bindgen@integration/embedded-rt-no-grow + cabi-realloc-extern, added a reference TCB arena (crates/gale-app-demo/tcb/cabi_arena.c defining __cabi_arena_realloc, trap-on-exhaustion), and reworked dissolve.sh:

gale-kiln     : 972 B native .text, 0 memory.grow   (was ~24848 B)
gale-app-demo : 316 B native .text, 0 memory.grow   (was ~24162 B)

~25× leaner, single-address-space, grow-free — the MCU library-OS object. wasip2 hosted demo unaffected (run.sh still run-demo()=53; feature gated not(target_env=p2)). rivet validate PASS.

Staying open until the upstream branch merges and we re-pin to a tag (currently an unmerged branch). The meld --memory shared component path remains separately gated on wit-bindgen#6 (component new rejects the env arena import) — not needed for this core-dissolve path.

[avrabe]

Cross-repo chain for the lean MCU dissolve — now complete & filed:

There are two routes to a grow-free, componentizable core (the precondition for meld fuse --memory shared):

• Route A — elide: wit-bindgen#6 (no cabi_realloc/memory.grow for scalar-only worlds) → component new works as-is. Best for our scalar gale:kernel.
• Route B — extern arena: wit-bindgen#4 (cabi-realloc-extern → env::__cabi_arena_realloc) + wasm-tools#2 (component new must accept that embedder-provided import). For components that genuinely need a bounded realloc.

Either route → grow-free component → meld#298/#300 --memory shared fuse → loom/synth dissolve → synth#383 shrink → run on the gust/kiln stack. gale#89 stays the consumer tracker; clears when one route lands end-to-end and a meld-fused grow-free core boots on the 8 KB part.

[avrabe]

Lean MCU pipeline PROVEN end-to-end (Route B) ✅

With pulseengine/wasm-tools feat/import-passthrough (the --import-passthrough fix, wasm-tools#2):

CM components (gale-app-demo imports gale:kernel; gale-kiln provides it over verified gale::*)
  → component new --import-passthrough env::__cabi_arena_realloc   (grow-free, validate OK)
  → meld fuse --memory shared --address-rebase                     (ONE merged memory, grow=0)
  → loom inline → synth cortex-m4f --relocatable                   (exit 0, 0 cross-memory skips)
  → 1328 B .text single-address-space native core

This is the full "drive components → meld down into the module → dissolve → MCU" path — a 1.3 KB fused native core from a Component-Model composition. Remaining to close gale#89: wire the fused core onto a gust-style TCB and boot it (qemu → Renode hermetic → physical F100 via REFLASH.md). Route A (wit-bindgen#6 scalar-elide, no flag) remains the parallel no-flag option.