LithosAnanake/CHANGELOG.md

Changelog

Milestone-level achievements only — not a commit log. For full history see git log on the relevant branch. Two products share this file: StarForth (hosted FORTH-79 VM, master) and LithosAnanke (bare-metal UEFI kernel, lithosananke). Entries are tagged by which branch(es) they landed on. Post-split note (2026-08-18): the combined monorepo this convention describes has been split into two separate repositories; LithosAnanke now lives here with master as its sole production line (no lithosananke branch in this repo). Entries below this note carry no branch tag for that reason.

2026-08-18 — Item 4.6 (Artemis Stadium migration) quota-ordering bug fixed; aarch64 BYE crash fixed; full 3-arch DoE campaign clean

Artemis's 30-rep surface stress campaign (item 4.6, migrating her block-heat arena to Stadium admission-on-allocate) was failing 100% of trials on all three architectures. Root-caused to capsule_birth.c: stadium_grant_quota() ran after IDENTITY exec, but Artemis's init.4th auto-runs the stress campaign as part of that same IDENTITY exec, so every STADIUM-ADMIT call hit a nonexistent quota slot and refused unconditionally. Moved the grant before IDENTITY exec. Verified 30/30 reps PASS on amd64, aarch64, and riscv64 (was 30/30 FAIL on all three).

Separately fixed a long-standing aarch64-only crash on BYE (present since at least 2026-08-08): arch_cold_reset() issued PSCI SYSTEM_RESET via smc #0, but this QEMU/AAVMF boot configuration has no genuine EL3/TrustZone secure monitor to answer an SMC — PSCI here is served via HVC instead. The instruction was trapping into the kernel's own exception handler on every boot, not corrupting memory or racing anything, despite investigation initially suspecting both. Root-caused via live gdb single-stepping, which also surfaced and fixed a real debugging-environment gotcha along the way: starkernel_kernel.elf is not the binary that actually runs under MONOLITHIC_BUILD (starkernel_loader.efi is, a completely separately-linked artifact) — every earlier breakpoint attempt had been aimed at the wrong file. Fixed: smc #0 → hvc #0. First clean aarch64 BYE exit recorded.

Also fixed, found incidentally during the aarch64 investigation: PSCI SYSTEM_RESET's function ID used the SMC64 calling convention (0xC4000009), which isn't a real PSCI function ID (SYSTEM_RESET has no SMC64 variant) — corrected to the valid SMC32 encoding (0x84000009).

Full three-arch DoE campaign (L8-DOE, 48 trials each) re-run clean on all three architectures post-fix: stress campaigns 30/30 PASS, clean BYE exit, no exceptions. Full investigation trail, evidence, and dead ends in FABRIC-2.md Sections H and I.

2026-07-24 — Kernel 3-arch acceptance test re-run; amd64 dict_hash non-determinism root-caused (lithosananke)

Re-ran the standing kernel acceptance test (ARCH={amd64,aarch64,riscv64} clean qemu) after this session's changes. All three booted cleanly to ok> with PARITY:OK and Hera/Hermes/Artemis E2E PASS, but amd64's dict_hash didn't match aarch64/riscv64 (which matched each other and the established reference exactly). Root-caused rather than dismissed: capsule_dict_hash_hook() (src/starkernel/capsule/capsule_vm_hooks.c) folds each word's execution_heat into the hash, and amd64's own boot log says its timer falls back to "RELATIVE mode (no determinism guarantees)" under this QEMU/hypervisor setup (no invariant TSC) — aarch64/riscv64 get proper deterministic hardware timers. Heartbeat tick count during boot, and therefore dict_hash, can legitimately vary run-to-run on amd64 from host scheduling jitter alone. Pre-existing, self-documented in the boot log, consistent with already-recorded ACL-RWT campaign data (amd64's baseline tick count was already distinct from aarch64/riscv64's). Not a regression from any change this session. Logs and DoE run data committed as usual.

2026-07-24 — amd64 ISR .hidden fix extended to aarch64/riscv64; hosted blk_ram sizing bug fixed (lithosananke)

Surveyed both branches for ASM files and C99 fallback layers on request; confirmed via SHA-256 that all six ASM-optimization headers plus arch_detect.h are byte-identical between master and lithosananke, but found amd64/isr.S had genuinely diverged — lithosananke carries a real fix (.hidden directives preventing GOT-indirection corruption of the runtime ASLR-relocation-delta computation) that master never received. Ported it there (see master's entry below). Investigated whether the same GOT-indirection bug class applies to aarch64/riscv64: it doesn't structurally (interrupts.c on those architectures only calls the vector-install function via PC-relative branches; nothing reads an asm-defined symbol's address from C) — added .hidden anyway as defensive hardening on request, working around a real ELF-vs-PE/COFF assembler incompatibility for the aarch64 loader (.hidden guarded with #ifndef _WIN32, since that loader is assembled via clang-18 --target=aarch64-pc-windows-msvc). Verified with full three-arch QEMU acceptance.

Separately, discovered and fixed (with explicit sign-off) a severe pre-existing bug: the hosted StarForth VM could not start at all on lithosananke, on any architecture. src/main.c's block-I/O RAM buffer was hardcoded to the old 1024*1024-byte size, stale relative to a deliberate BLK_RAM_BLOCKS=2080 block-subsystem redesign (moving the RAM/ramdrive boundary to user-visible LBN 2048) that the kernel-side bootstrap already correctly followed. Fixed to derive the buffer size from the same shared constants; verified across all three hosted architectures via QEMU (native amd64, qemu-aarch64-static, qemu-riscv64-static).

2026-07-24 — pgo build option removed; hosted 3-arch acceptance test documented (both)

Removed the TARGET=pgo profile and its associated pgo/pgo-build/ pgo-perf/pgo-valgrind/bench-compare targets from the hosted Makefile on both branches, plus the now-orphaned scripts/pgo-workload.sh. Verified make TARGET=pgo now errors cleanly on each branch.

Also proved out and documented, for the first time, that the hosted VM starts cleanly not just on amd64 but on aarch64 and riscv64 too (cross-built, run under qemu-{aarch64,riscv64}-static user-mode emulation) — docs/lithosananke/hosted-acceptance-test/README.md. This is what surfaced the blk_ram sizing bug above.

2026-07-07 – 2026-07-08 — Kconfig build-configuration system (both, committed on lithosananke)

A Linux-kernel-style Kconfig system replaces the two Makefiles' scattered, drifting ?= knobs with one discoverable, optional configuration tree. Vendored the real scripts/kconfig tooling (conf/mconf/qconf — headless, ncurses, and Qt frontends) into tools/kconfig/, wrote a ~40-symbol tree (Kconfig.arch/variant/physics/heartbeat/kernel), and bridged it into both Makefiles via mk/Kconfig.mk. Entirely opt-in: nothing changes for a plain make until a developer generates a per-architecture build/$(ARCH)/.config, and command-line overrides always win regardless. Found and reported two real, previously-invisible drift bugs along the way (TRANSITION_WINDOW_SIZE diverging live in kernel builds, MINIMUM_PREFETCH_ROI off by an order of magnitude in its Q48.16 encoding); the latter was fixed on request. Documented in docs/working/architecture/getting-started/DEVELOPER.md.

2026-07-07 — Artemis/riscv64 determinism gap closed (lithosananke)

Root-caused a dict_hash divergence on riscv64 that had been mis-attributed to an architecture quirk: Loop #3 heat decay was reading real wall-clock nanoseconds instead of the heartbeat's execution-driven tick counter, making it sensitive to a riscv64-specific I/O busy-wait. Converted decay to be purely tick-based across all six call sites touching it. All four VM identities (Hera, Artemis, Hermes ×2, Mama) now byte-identical across amd64/aarch64/riscv64.

2026-07-05 – 2026-07-07 — VM Fleet Attractor experiment concluded (lithosananke)

Multi-week investigation into giving the L8 Jacquard mode selector a real per-VM physics channel (dynamic fleet heat replacing hardcoded compudynamics.4th/fleet-k.4th constants) reached a validated end state: a 180-run, three-architecture DoE campaign found the self-tuning inference loop was starved by a shared-counter bug, replaced a log-linear regression estimator with direct rate recovery once fixed, and closed with L8 wired into a real causal channel into fleet tuning — the pathway earlier iterations had found architecturally missing. Three-arch acceptance confirmed throughout.

2026-07-04 — K-PUSH dictionary-shadowing corruption fixed (lithosananke)

Multi-day trap-and-debug campaign closed: FORTH-79 word shadowing broke under the dictionary's heat-based reorganization, corrupting K-PUSH. Root-caused and fixed; three-arch confirmed.

2026-07-03 — Artemis BAM storage subsystem, Phase 1 (lithosananke)

Block Allocation Map storage for the Artemis VM accepted: Q48.16 block heat tracking, ART-COOL/ART-REAP reclamation words.

2026-06-27 – 2026-06-30 — Hermes inter-VM messaging complete (lithosananke)

17-block message/channel layer between VMs finished and verified: 16/16 inter-VM smoke tests passing, three-arch. Follow-on "completion steps" batch (async delivery, channel negotiation, K-FLEET integration) landed the same week. INPUT_BUFFER_SIZE restored to 1025 on this branch with the correct line-vs-block-buffer rationale (see note below on master's divergence).

2026-06-23 – 2026-06-25 — Tripod (Hermes/Artemis) scaffold stood up (lithosananke)

Named multi-VM roles (Hermes, Artemis) with dedicated prompts and CONNECT-HERMES/CONNECT-ARTEMIS wiring; verified booting live, pre-REPL, on all three architectures. Compudynamics — the physics-runtime tuning discipline — ported into FORTH itself via VM-EXEC/VM-STEP primitives, the first self-hosted physics orchestration. K≡1.0 conservation wired end-to-end (Phase 6+7) with Artemis block storage passing PHASE6-TEST green.

2026-06-17 — LaTeX publication system completed (master)

Three-volume formal documentation system finished: 11 publications, 209 source scraps at 100% coverage, clean build. StarForth version bumped to v3.1.0 / LithosAnanke to v1.5.3 in the same window.

2026-06-16 — LithosAnanke v1.5.1-FINAL — framebuffer/VT100 console milestone (lithosananke)

Closed the framebuffer terminal milestone; three-arch re-acceptance. ACL-RWT DoE campaign confirmed the same day as the first genuinely ACL-active campaign (an interpreter-hook bug that had silently bypassed ACL checks in every prior run was fixed just before this one) — all nine cells of a 3×3 Latin square (seeds × architectures), +0.0054%–+0.0088% overhead, CV = 0.000%. Report finalized for patent support material.

2026-06-13 – 2026-06-14 — Word-level ACL system, Phases 1–7 complete (both)

C infrastructure (acl_ttl/acl_allow/acl_mode/acl_pinned fields on every dictionary entry), the ACL.4th FORTH policy layer, the zuse bootstrap superuser, Isabelle/HOL proof stubs, and LithosAnanke kernel parity all landed — all three kernel architectures boot to zuse)ok>. Phase 8 (Ed25519 PKI / thumbdrive challenge-response) is the only remaining item on the ACL roadmap.

2026-06-11 – 2026-06-12 — Bare-metal ACL baseline DoE campaign (lithosananke)

3×3 Latin-square baseline campaign (seeds × architectures, 30 replicates each) completed across all nine cells. Analysis report drafted, revised through a peer-review pass, and prepared for SSRN submission.

---

Known issue surfaced while compiling this changelog, not yet resolved: master's include/vm.h currently defines INPUT_BUFFER_SIZE as 256, violating the project's own documented invariant (it must be 1025 — LOAD copies up to 1024 bytes of block content through the same vm_interpret() path interactive REPL lines use, so anything past byte 255 is silently dropped under the 256 cap). lithosananke has the correct value. Not fixed as part of this changelog/README pass; flagged for a future session.