Reviewed: 2026-10-30 at the M2 exit review, against the M2 branch stack ending in docs/m2-11-exit; previously 2026-10-14 against baseline f0f3afb473a61ebf02a47a9106cf9690f98ece2e. This inventory is maintained manually; the requirement matrix generated from executed cases complements it and is not a published support matrix.
This page is the authoritative inventory of implementation and available evidence. Specifications define target contracts; their acceptance dates do not mean that all described APIs exist. The roadmap determines release eligibility. Timing and evidence defines how a claim becomes substantiated. Port pages explain implementation details and historical measurements.
| Area | Current implementation | Evidence and limits |
|---|---|---|
| Execution and scheduling | UP fixed priority, base bitmap/FIFO scheduler, tiny thread table, scheduler lock, thread lifecycle; idle context unreachable through handles, and slim (control-block prefix only) on tiny without an idle thread (ADR-050); optional round-robin time slicing below a threshold level (A63, off by default) | Common conformance cases; tiny has unique base priorities and restricted scheduling semantics; slicing is list-scheduler only, bridge-checked with the periodic clock and conformance-tested in both clock modes on native, not measured on hardware |
| Waiting and time | Shared wait protocol, sleep, deadlines, cancellation, suspend overlay, ordered timeouts | Native virtual time exercises periodic and tickless modes; embedded ports currently reject tickless |
| Synchronization | Semaphores, mutexes with inheritance, ceiling, recursion and owner-death policy; event flags with ANY/ALL/CLEAR waits and binding (M2-05); condition variables with the mandatory relock (M2-06) | Conformance and native differential bridge; inheritance and deadlock detection have a configured depth limit |
| Communication | Copy-by-value message queues: direct hand-off to a waiting receiver, slot hand-off from a waiting sender, send to front, peek, EMB_NO_WAIT calls from ISRs and a receive-side binding (M2-07) |
Conformance and native differential bridge; no pipes, buffer pools or zero-copy ownership queues; the masked copy time is bounded by configuration, not measured on hardware |
| Software timers | One-shot and drift-free periodic timers in their own deadline list (ADR-047), with callbacks on a work queue or in interrupt context, closed-form catch-up and overrun counts, stop states and stop_sync (M2-09) |
Conformance and native differential bridge; list scheduler only (not tiny); arming overhead and jitter not measured on hardware |
| Deferred work | Work queues on a kernel thread with FIFO execution, idempotent and ISR-safe submit, resubmission while running, cancel and synchronous cancel, flush, destroy joining the thread, a system queue linked only when referenced (M2-08, ADR-048), and delayed work items over internal timers (M2-09) | Conformance and native differential bridge; list scheduler only (not tiny); optional per-queue statistics (A65), conformance-tested on native and spot-checked on QEMU Cortex-M4 |
| Notifications | Per-thread bits; bindings from semaphores, event flags and message queues to a thread, with unbind; a bound thread cannot be destroyed (ADR-046) | Binding takes a one-bit mask; pipes and pools do not exist to bind; restart generations need the isolated profile |
| Objects | Caller-owned generated storage and POINTER handles; per-type active-object lists for every handle object type (CONFIG_EMB_OBJ_LIST) used by init to refuse live storage (ADR-045) |
No INDEXED or TABLE handles, capability rights, partition restart, system freeze, emb_obj_foreach() or debug-descriptor export |
| Native port | POSIX host-thread gate, deterministic IRQ injection, virtual time | Base/tiny development workflow; no current wall-clock, asynchronous signal stress, simulated isolated profile or Win32 backend |
| Embedded ports | ATmega328P; Cortex-M0/M3/M4F/M7/M33; RV32IMAC and RV32IMC machine mode; Xtensa LX6 (the classic ESP32's PRO_CPU, one core, level-1 interrupts, ADR-054); boards Arduino Uno, NUCLEO-F446RE, ESP32-C3-DevKitM-1, ESP32 DevKitC, Raspberry Pi Pico 2 W (RP2350, built only), QEMU's MPS2 (AN385, AN386, AN500, AN505), micro:bit and RV32 virt; the NUCLEO-F446ZE from its description alone; a project's own boards outside the tree through EMB_BOARD_ROOT (A76) |
QEMU jobs and EmbSIM jobs (Uno, NUCLEO-F446RE, NUCLEO-F446ZE, ESP32-C3-DevKitM-1) defined in CI; a HIL runner and presets (avr-uno-hil, cm4-nucleo-hil, procedure) exist; the first hardware run, on an Arduino Uno, passed every test in the checked (17), release (15) and T1 (12) builds, and the second passed checked 17/17, release 17/17 and T1 13/13 with 53 misuse checks executed on the part (A89), its console logs and image hashes filed in engineering/hil-runs/. The classic ESP32's bring-up probe ran on a physical board (report), and then the kernel's Xtensa port, built by EmbCC, passed 21 of 21 tests there, 157 cases and 74 fault checks that reset the board (report); CI builds the port with Espressif's GCC and Clang, and both builds passed the suite on the board too. No Cortex-M or RISC-V board has run on hardware yet |
| Configuration/build | Kconfig subset, port manifests, CMake/Ninja, layout probe, footprint tool; toolchain neutrality (ADR-053: one port per architecture, GNU extensions confined to arch/ and include/emb/compiler/) |
GCC/Clang native CI and cross-GCC QEMU jobs; an EmbCC-kernel/GCC-host native CI leg on a pinned release; ten legacy EmbCC capabilities reviewed with scoped local checks, but embedded EmbCC builds and compiler qualification remain incomplete |
| Observability | M1 trace sink and fault reporting; trace records for every outcome of the M2 waits and for queue, work and timer events (IDs 31 to 44); a crash record written before the fault action and kept across reset, with emb_crash_record_get()/_clear(); the debug descriptor, which GDB reads to list threads and walk a switched-out thread's stack (A107) |
The record has no registers or flight recorder yet; CTF rings/metadata, deferred log decoder and worst-case monitors remain planned. The descriptor has been read on one board, the NUCLEO-F446RE through OpenOCD, besides QEMU, and does not yet describe the AVR and Xtensa frames |
| Documentation | Sphinx/MyST manuals and native tutorials, Doxygen/Breathe C API, Sphinx Needs design catalog, JSON/CSV downloads and annotation audit | Strict local HTML build, complete extracted function annotations, source-preservation and catalog-integrity checks; merged 37887c9 Pages deployment and served evidence-workflow page independently observed successful |
| Repository process | Issue forms, PR checklist, changelog, release/M2 work checklists, stable CI merge checks, authored-commit DCO checker and a main ruleset policy | Local syntax and DCO/gate failure checks; GitHub's effective main rules report PRs, resolved conversations, the three strict CI checks and deletion/non-fast-forward protection; no remote rejection experiment claimed |
| Semantic model | UP scheduler/wait/mutex/notification/lifecycle scenarios; binding lifetime, event flags, condition variables, message queues (modeled before their implementation, KRN-IPC-022), work queues, software timers, delayed work and combined scenarios; exhaustive section-level exploration and randomized walks | Scenario-scoped exploration, not a proof of all C execution, arbitrary programs, SMP or hardware protection |
Barriers, pipes and buffer pools; indexed handles and capability tables; static partitions, syscall marshallers and supervisor restart; execution budgets and partition shares; the system description and most generated outputs (the hardware generator validates SoC and board descriptions and generates a first set of outputs for the STM32F446 and ATmega328P, from which every board but the ESP32-C3 builds: NUCLEO-F446RE, NUCLEO-F446ZE, Arduino Uno, MPS2 AN385, AN386 and AN505, RV32 virt); importers beyond the SVD, Zephyr DeviceTree and STM32 pin ones (non-STM32 pins and DeviceTree boards); reusable driver classes; power management; boot/update services; full observability formats; a published support matrix for releases (a board support matrix of tiers from evidence exists; executed-case requirement traceability and kernel statement/branch coverage exist, see case evidence and the M2-10 campaign); fuzz/HIL campaigns; SMP/AMP ports and transports.
An API or path named in a design or a requirement's verification field may be a planned deliverable. Check this inventory and the filesystem before treating it as available.
The M2-10 integrated campaign, source f563ef5, passed 158,800 differential runs across seeds 1–10 on native checked/clang/periodic/release/tiny with no divergence; all 21 local configurations (native, sanitizers, coverage, QEMU, AVR, EmbSIM) passed; kernel coverage was 91.4% of lines and 72.5% of branches. Hardware-in-the-loop and physical timing remain unrun.
The reference controller, examples/controller at dc91e59, ran on a NUCLEO-F446RE on 2026-11-07. It passed 9 of 9 phases: overload into degraded mode and back, 1 kHz, a 30-minute soak with no miss or drop (later an overnight soak of 10 hours 6 minutes, 3,636,800 samples, also with none: soak report), an injected fault that the crash record decoded to its source line, and a hang reset by the watchdog. A logic analyzer then measured its latency from outside: from the ADC's interrupt to the control thread at work, 9.50 µs at the median and at most 10.0 µs over 24,998 samples, at 100 Hz and 1 kHz and in degraded mode (timing report).
The onboarding runs on the baseline used GCC 14.2.0, Python 3.12.14, CMake 3.31.10 and Ninja 1.13.0 on Linux. native-gcc, native-release, and native-tiny each passed ten CTest targets, including nine conformance executables and the differential bridge. Those are runner targets, not ten individual assertions. The base checked run reported 82 internal conformance cases with zero skips, and 210/210 bridge runs agreed with the model with one scenario excluded for that configuration. Counts for other configurations must be read from their own logs.
The reference model passed 44 pytest tests with Hypothesis installed. The banned-construct check reported zero findings. Configuration generation for native/base passed. A repeat setup run reused build outputs and passed. These are current-workspace observations, not archived release artifacts or evidence of a new machine, compiler matrix, physical board, or published environment.
The M2-01 timer correction, source ca0ed3f, subsequently passed 53 model tests, eleven CTest targets on each of native checked/release/tiny/periodic, sixteen exact live regression inputs and 11,200 unfiltered deep-campaign runs with zero divergences. Its source/build/configuration identities and outputs are archived separately from the original failing captures. The corrected model accepts all seven original archived traces; production kernel and architecture sources are unchanged.
The differential evidence workflow, source 8d7bca4, adds automatic failure/probe bundles, offline/live replay and bounded confirmed schedule reduction, including full resolved-configuration checks independent of checkout paths. It passed 67 model/tooling tests, twelve CTest targets per checked/release/tiny/periodic native configuration and another 11,200 deep-campaign runs with zero divergences. Its workflow checks deliberately create search-budget failures to validate capture/replay, and keep those separate from actual kernel evidence. Native CI is configured to retain default failure bundles on failed jobs.
The bounded seed-5 triage, current source 37887c9, subsequently passed 56,000 unfiltered executions across seeds 1–10 on those four presets, with the existing applicability exclusions. Rebuilding the source that first recorded the report reproduced the expected old seed-3 mismatch (1399/1400 agree) and passed 4,200 seed-5 runs on checked/release/tiny. The historical cancellation report was subsequently closed as not reproducible (input unavailable, cause unknown; maintainer decision 2026-10-27), not fixed and not classified. Commands, bounds, source/build/configuration identities, logs and exact current cancellation schedules are archived separately.
The EmbCC review, compiler source a3c3ef4 and RTOS source 2e59c2a, confirms upstream implementation of the ten former compiler gaps within their stated restrictions. Ten focused local checks passed; AVR linking/program-memory fixtures and small RV32/AVR atomic values executed under EmbSim. No local QEMU, board, live-debug, CMSE-crossing or complete EmbCC RTOS conformance run is claimed. The host compiler bootstrap required local flags and was not warning-clean. Linux executable configure lacked a built target runtime; compile-only configure succeeded but the RTOS storage probe failed on the missing EmbCC EMB_LAYOUT_MARKER. Upstream PR #119 reports RTOS experiments with an injected marker, recorded separately from current CI qualification.
.github/workflows/ci.yml defines native GCC/Clang, release/tiny/periodic, ASan/UBSan and TSan jobs; Cortex-M, RV32 and AVR cross-GCC/QEMU jobs; the AVR, Cortex-M and RISC-V conformance suites on EmbSIM's Uno, NUCLEO-F446RE and ESP32-C3-DevKitM-1; formatting, banned-construct and configuration checks; model tests; and AVR footprint gates. A job definition is not a successful run. Link immutable run results and artifacts before advertising a target/version as supported or validated.
On 2026-10-14, GitHub reported the complete pre-workflow-preparation matrix successful for merged source efff40cc2dbf803a56bab5c3fc844b78a4837f9f: CI run, docs build and Pages deployment. The configured site URL is https://emblinkos.github.io/EmbLinkRTOS/. These are observed workflow/job outcomes, not individual requirement coverage or physical hardware qualification; the new stable checks need their own PR results.
On the same date, the manual Documentation build and Pages deployment succeeded for reviewed main source 37887c9e92e2242ff4d11cfa58c6972fce4b0e0e. An HTTP 200 request to the served evidence-workflow page confirmed that snapshot's content. This later publication includes the correction/capture tooling, not this subsequent unmerged triage/contract-review documentation. GitHub's effective branch-rules API also reported the three strict required checks (CI merge gate, Documentation build, Repository policy), PRs with zero required approvals, resolved review conversations, and blocked deletion/non-fast-forward updates on main. Observing settings is distinct from testing a rejected push.
The later main publication succeeded for 2e59c2a37b36c90e5f15892fb258c8a2e21be306, and the served M2-02 contract-review page returned HTTP 200 with its expected content. This publishes the merged seed-5 triage/contract proposals, before the present EmbCC review update.
The repository preparation adds a stable CI merge gate depending on all of those matrix/lint jobs, rejects empty CTest runs, and selects Bash with pipeline failure propagation for footprint commands. Documentation build names the strict docs check and Repository policy checks matching author DCO trailers on non-merge commits. The repository workflow describes the exact scope, limitations and additive main ruleset policy; GitHub settings must be activated and verified separately. Local validation used actionlint 1.7.7, parsed the issue forms/policy, exercised success/failure/skip/cancellation/empty dependency outcomes, and checked DCO rejection/acceptance with an actual temporary Git range. These checks do not establish remote CI or live branch enforcement.
Since 2026-11-08, clang-tidy 18 and cppcheck 2.22.0 run over the product code in 14 configurations with 0 findings (static analysis), and CI fails when a configuration loses a requirement's executed evidence or kernel coverage falls below its floors (92.5% of lines, 70.7% of branches): case evidence. The double build of BLD-004 is a gate since then too: the same commit from a second checkout gives byte-identical images on native, Cortex-M (GCC and Clang), RV32, AVR, the Pico 2 W and the reference controller (reproducible builds). cppcheck's MISRA addon runs against a baseline that may only shrink. Generated requirement completeness (every in-scope requirement covered), MISRA compliance, include-boundary checks, fuzzing and HIL are not currently CI gates. The roadmap must deliver them for the released scope rather than assume the engineering-system target description is already enforced.
The separate documentation toolchain is installed in .venv-docs with a checksummed Doxygen 1.15.0 binary under build/tools/. Its local HTML build extracts 361 design requirements and 82 current API functions. All extracted functions have eight contract tags and parameter descriptions; their declared requirement IDs resolve in the normative catalog. The strict build rejects omissions and unresolved IDs. This is extraction/documentation evidence, not API semantic compliance or test coverage. A related requirement can describe more than the implemented subset. The documentation integrity suite checks malformed/duplicate/empty catalogs, annotation omissions and wrapped contract fields, and keeps design acceptance separate from execution evidence.
The user-guide pass also configured native/base with stack statistics and two TLS slots in a separate native-inspect build; all ten CTest targets passed. Both complete tutorial programs were compiled with strict C11 warnings and run with the documented expected output and zero exit status against checked base, release, tiny and this diagnostic configuration. The diagnostic callback fragment compiled as C11, and all three fragments compiled as C++17 against native/base headers. These are local tutorial/configuration checks; they add no physical-board or universal API-coverage claim.
The EmbCC layout-marker correction, source 3578006, supplies the portability macro; storage/assertion/kernel-archive compilation passed on eight configurations. The subsequent Linux native integration, source df32ec2, adds explicit experimental EmbCC-kernel/GCC-host presets and a build-time layout guard. Actual checked/release/tiny/periodic mixed builds passed all twelve CTest targets each and 11,200 seeded model comparisons in total, with zero divergences. GCC compiles the POSIX port, board, applications and harness and links the real Linux runtime; this is not an all-EmbCC host build. The latter still fails on glibc header compatibility after an explicit system-header/link experiment. CI now runs the four mixed presets with the pinned EmbCC 0.0.1-beta release on every pull request (merge-gated; 20, 20, 20 and 17 CTest targets passed on the first run); deeper campaigns remain scoped local evidence.
- Hardware timing and total memory use are demonstrated for one application, the reference controller on the NUCLEO-F446RE.
- Its latency was measured with a logic analyzer (timing).
- Its flash and RAM are accounted for byte by byte, with a static stack bound for each thread and the interrupt stack that CI checks (memory).
- Other applications need their own roots file for the bound. The kernel-only footprint subtotals still exclude port, board, runtime and stack costs.
- The Cortex-M base TCB figures in the port page exceed R-003 T4's target; that is a disclosed target miss, not a failed native setup.
- The M2-01 timer correction addresses the reproduced seed-3 model timing gap, and the evidence workflow provides automatic bundles, exact replay and bounded schedule reduction. The historical seed-5 cancellation report, whose original schedule and build are unavailable, is closed as not reproducible: input unavailable, cause unknown after the bounded triage. It remains a known limit: passing current campaigns does not establish its cause, and the triage names the evidence that would reopen it.
- AVR launch/restore uses
out SREG; pop r0; ret; pending IRQ coherence during that tail remains a physical validation obligation (A27), not a verified atomic out/ret pair. - Fault-expectation cases run on native (fork) and, since A89, on every board whose reset keeps RAM: QEMU's MPS2 machines and RV32 virt, EmbSIM's Uno and NUCLEOs, the physical Uno, where 53 ran, the physical ESP32, where 74 ran, the physical NUCLEO-F446RE, where 74 ran with GCC and with Clang, and the physical Pico 2 W, where 75 ran with each (2026-11-07). They are still skipped on QEMU's Uno and micro:bit and on EmbSIM's ESP32-C3. No RISC-V silicon has run them.
- The M2-02 IRQ correction implements larger-is-more-urgent in the advertised kernel-aware domain: Cortex-M derives the range/encoding from board priority width and BASEPRI; native, AVR and current RV32 expose level 0 only. Invalid requests fault when checked or return an error in release before controller writes. SysTick defaults to highest kernel-aware urgency. Initial QEMU M3/M4 and F446 EmbSIM runs passed the new register/order/preemption cases; local checked/release/tiny/periodic native and experimental EmbCC-kernel/GCC-host checked builds each passed 13 CTest targets. Physical latency/jitter, raw kernel-independent vectors and generic hardware-description generation remain unqualified/unimplemented; the raw class is refused at configuration.
- Thread diagnostic indexes are now independent of priority in every scheduler profile, unique among live objects and reserved through termination until successful destroy. A bounded bitmap replaces base's saturating counter. See the identity report for admission limits, costs and verification. Public API misuse faults now carry stable numeric entry IDs, including shared-helper paths; generic/manual faults keep ID zero. See the fault attribution report. Retained crash records and full observability remain incomplete.
- Thread init refuses storage that still holds an undestroyed thread (fault when checked,
EMB_EBUSYin release): exact with the all-threads list, otherwise from the type tag, active state and reserved identity. Semaphore and mutex init refuse storage holding an active object the same way, exactly, withCONFIG_EMB_OBJ_LIST(default except tiny); on tiny their double init remains undefined. See the object lifecycle report. Diagnostic ID reuse after destroy does not add generation checks to POINTER handles; storage lifetime and external synchronization remain caller obligations. The M2-02 review records the remaining lifecycle proposals. - The former ten-item EmbCC compiler backlog is implemented upstream and the RTOS layout marker is supplied. An explicit mixed Linux native strategy now executes the kernel; all-EmbCC POSIX host compatibility, AVR flash/registration macros and compiler-matrix CI/qualification remain open. Available CMSE, atomics, DWARF and embedded C++ retain their documented scope and restrictions; see the capability review and marker correction.
- Planned isolation must meet SPEC-010's DMA, pointer lifetime, generation and restart obligations before fault containment is claimed. CPU protection alone is insufficient.
Record source revision, configuration, tool versions, runner results, individual cases, skips and artifacts together. Move a feature from specified to implemented after inspecting the code; call a requirement verified only after its named evidence passes for the applicable profile. Do not infer universal coverage from one representative configuration. Preserve historical port measurements with their original toolchain and feature set.