Expose cross-component value ranges + bounded-memory facts from the fused core for downstream specialization

[avrabe]

Part of the meld→loom→synth 0.70× (beat-LLVM) roadmap (gale benches/gust/optimization/BEAT-LLVM-0.7x.md (PR pulseengine/gale#119)). After meld fuses the components into one merged-memory core, the whole-program call graph and value flow are known (and memory is bounded, 0 memory.grow) — information no per-TU compiler (LLVM) has.

Ask: emit, as metadata on the fused module for loom/synth to consume, the facts fusion makes available — cross-component value ranges, constant call-site arguments, dead parameters, and the bounded-memory guarantee. These become optimization premises enabling cross-component specialization (constant-arg propagation, dead-arm elimination, range-driven check elision) that LLVM can't do across translation units. Pairs with the loom issue (carries the facts in IR) and the synth issue (consumes them to specialize). The payoff is dissolved code faster than native LLVM because fusion + proof expose facts the native build never sees.

[avrabe]

Design note — emit fusion premises onto the existing scry boundary, don't reimplement the analysis

Scoping this against the toolchain as it actually stands:

The value-range / constant-arg / dead-param analysis is scry's job, not meld's. scry is the sound abstract interpreter (scry-interval = interval domain, scry-octagon, scry-analyze-core already produces reachable_from_exports, intervals, stack-usage, call-graph) and is consumed by synth. meld reimplementing dataflow would duplicate scry-interval and wouldn't be sound-by-construction. So meld's contribution to the "beat-LLVM 0.70×" premise set is the facts only fusion knows — the ones that tighten scry's fixpoint.

These belong on the existing meld→scry boundary, not a new section. meld already emits the versioned component-provenance custom section (#192/FEAT-008; provenance.rs), whose format is owned by scry-provenance per DD-002 (meld = producer, scry = consumer; fused_module_sha256-bound; v2 code_range; "adapter provenance = a future v2 ask"). #313's facts are the next premise fields on that boundary.

Proposed inc 1 — fusion premises (cheap, fusion-unique, sound)

Extend component-provenance (vN+1) with:

• bounded_memory — no memory.grow in the fused core (meld already detects this: memory_probe::module_uses_memory_grow; it's the --memory auto premise). Lets scry assume a fixed linear memory / drop grow-reachability widening.
• closed_world — all cross-component imports internalized (meld knows this: stats.imports_resolved / no residual cross-component imports). Lets scry tighten reachable_from_exports and assume no external mutation of fused state between components.

These are emit-only on meld's side (no new analysis), and they're premises scry can't soundly assume on its own without the fusion guarantee. value-range/const-arg/dead-param then fall out of scry's interval/reachability fixpoint given these premises — fed to synth.

Coordination (format owned by scry-provenance, per DD-002)

• scry-provenance: add the premise fields to the component-provenance decoder (vN+1); a missing field stays backward-compatible (older meld → premises absent → scry falls back to conservative).
• scry-analyze-core: consume the premises as analysis assumptions (relates to scry#5 interval fixpoint, scry#51 sound call-graph/reachability).
• meld: emit the fields. I can land the meld producer side as soon as scry-provenance fixes the vN+1 field layout (so meld doesn't emit a shape scry won't decode).

Out of scope for meld (stays in scry)

Interval/value-range, octagon, taint, the actual constant-arg/dead-param derivation — all scry-interval/scry-analyze-core. witness (MC/DC) coverage could additionally feed reachability/dead-branch facts later, but that's a separate source.

Proposing this as the inc-1 contract; happy to open the matching scry-provenance issue for the vN+1 field layout if that's the right split.

[avrabe]

Consumer-side format ask filed: pulseengine/scry#63 (component-provenance v3 premises block: bounded_memory, closed_world). meld's producer-side emission (this issue) lands against whatever v3 field layout scry#63 fixes — so meld never emits a shape scry won't decode. inc 1 is unblocked the moment scry#63's layout is agreed.

[avrabe]

Boundary is now live end-to-end. scry v2.0.0 shipped the SCPV v3 consumer (scry#63 / scry PR #64; scry-sai-provenance+scry-sai-core on crates.io), and meld's producer (#314) emits the same layout — 43-byte header (magic|ver=3|bounded|closed|sha256[32]|count) + length-prefixed entries — byte-for-byte to the ratified spec. The meld→scry component-provenance channel, dead since #192 (JSON producer vs binary decoder → BadMagic → provenance=None), now actually connects.

Notable convergence: scry's clean-room independently caught the same untrusted-entry_count DoS the Mythos gate flagged on meld's side — both decoders now bound the pre-allocation by remaining payload rather than the wire count. Two independent analyses, same defense.

Producer side (#314): green, CLEAN, Mythos NO FINDINGS after three real findings hardened it (the entry-count DoS + two closed_world over-assertions → closed_world is now the provably-sound zero-imports tautology). Awaiting maintainer review/merge (Tier-5). scry#63 closes when #314 lands. Premise consumption (bounded_memory → tighter regions, closed_world → complete reachability) tracked on scry#5/#51.

[avrabe]

Closing — meld's side is complete and shipped in v0.37.0 (#314).

The fused core now emits the fusion-unique premises in component-provenance (SCPV v3): bounded_memory (!uses(memory.grow)) and closed_world (provably-sound zero-imports). scry v2.0.0 consumes them byte-for-byte (scry#63); the meld↔scry boundary is live end-to-end.

The remaining part of the original ask — cross-component value ranges / constant args / dead params — is by design scry's abstract interpretation (scry-interval / scry-analyze-core), fed by these premises, not meld's to compute (SR-45). That work continues downstream in scry/loom/synth.

Nothing further for meld here. Reopen if the premise set needs extending.