uce/WASM-PROPOSAL.md

# WASM-PROPOSAL: WebAssembly Unit Runtime for UCE

- **Status:** design guide; Phase 0–4 spikes complete and gated (2026-06-12).
  Next: production implementation, starting with the Phase 3 work plan
  (item 1, real `uce_lib` core compile); the starter parity bar
  (`tests/run_network_tests.py --match starter`, 14 cases) is already green
  against the native backend.
- **Scope:** replace the native unit pipeline (generated C++ → clang → `.so` →
  `dlopen`) with per-unit WebAssembly modules executed in a per-request,
  runtime-linked workspace, exposing the same API surface to page code,
  enabling memory safety, better execution control, and paving the way for
  supporting more source languages in the future.

---

## 1. Motivation

Two long-standing structural problems and one strategic opportunity share a
single root cause: **request code shares an address space and an allocator
with the runtime.**

1. **The arena attempt failed for a structural reason.** The
   `GLOBAL_ARENA_ALLOCATOR` design in `_scratchpad.cpp` swapped the global
   `operator new`/`delete` against a `current_memory_arena`. A single global
   allocator cannot distinguish request-lifetime allocations from
   process-lifetime ones: during a request, server-lifetime structures
   (compile registry, sessions, config trees, unit statics) also allocate.
   Under the arena those dangle on reset; with `delete` as a no-op,
   system-allocated objects released mid-request leak. The lifetime
   distinction lives in the type system and call graph, not at the allocator
   boundary — fixing it in-process means threading PMR allocators through
   `String`, `DValue`, and every container.

2. **Fault recovery is best-effort, not sound.** The current
   SIGSEGV → `sigsetjmp`/`siglongjmp` recovery performs no unwinding, skips
   destructors, can resume over corrupted worker state, and (see
   RECOMMENDATIONS.md 1.5) cannot reliably produce a useful backtrace.
   A faulting unit *can* have scribbled on runtime memory before the signal.

3. **UCE can only ever host fully-trusted code.** A `.uce` page is arbitrary
   native code with full process privileges. Multi-tenant or user-supplied
   page hosting is structurally impossible in the native model.

A WASM execution model deletes the shared-fate fact itself rather than
patching around it:

- **Arena by construction:** each request runs in its own linear memory,
  dropped wholesale at request end. Request-lifetime memory physically cannot
  outlive the request; server-lifetime state physically cannot live inside
  it. The `_scratchpad.cpp` design becomes correct because the boundary is
  structural, not typological.
- **Sound recovery:** a trap (null deref, OOB, stack exhaustion) is a defined
  host-side error that unwinds cleanly, with a precise guest stack trace, and
  the unit cannot have touched host memory. Render error page, drop
  workspace, keep serving — actually correct, not hopeful.
- **Capability security:** page code gets exactly the imported API surface
  and nothing else. Multi-tenant hosting becomes possible.
- **Secondary wins:** architecture-independent unit artifacts (no clang
  required on prod), safe module unload/replace (vs. never-safe `dlclose`),
  first-class limits (linear-memory cap = RAM limit, epoch/fuel = CPU
  timeout), per-request memory stats for free.

**Expected outcome:** ~1.2–2× compute slowdown vs. native (still far ahead of
interpreted runtimes); a real ABI/membrane design; ownership of a custom
loader; toolchain rough edges (§10). More isolated, shared-nothing PHP architecture
but with a good story about websockets, generic sockets, background tasks,
and other long running processes.

---

## 2. Rejected alternatives (recorded so they stay rejected)

- **In-process PMR arena.** Requires re-typedefing `String`/`DValue` and
  threading allocators through the entire codebase; the failed global-swap
  shortcut is the only cheap version and it is unsound (§1.1).
- **One instance per component.** UCE components are function calls, not
  RPCs: callees receive the context **by reference**, mutate `context.call`,
  share the `ob_*` capture stack and `ONCE` dedup state. Per-component
  instances force serialize/copy/deserialize of the context on every
  `component()` call (a dozen+ per page in the starter), require
  host-mediation of the ob stack, and silently change reference semantics to
  copy semantics. Rejected.
- **One linked module per app ("the blob").** Introduces an "app" concept
  UCE does not have, makes every edit a global relink, turns the artifact
  cache into a build graph — webpack's worst traits without its benefits.
  Rejected. The file stays the unit.
- **Eager pre-loading of known units at worker warm-up.** Rejected; loading
  is strictly lazy, on first explicit call, preserving current semantics.

---

## 3. Architecture overview

### 3.1 Execution model

```
host process (linux_fastcgi, per worker)
│
├─ vendored wasm runtime (§10.1)
├─ unit artifact cache: one PIC .wasm per unit (replaces per-unit .so)
├─ loader (§6): dylink parsing, base allocation, GOT resolution,
│   symbol registry, ABI stamp check
├─ core snapshot: "core module, initialized" memory+table image
│
└─ per request: WORKSPACE
   ├─ linear memory (CoW-born from core snapshot)
   ├─ shared funcref table
   ├─ core module instance  (uce_lib + libc compiled to wasm)
   ├─ unit module instances (loaded lazily, on first call, incl. mid-request)
   └─ host handle table     (sqlite/mysql/file/socket handles + closers)
```

- **Workspace = request.** Born from the core snapshot, dropped at request
  end. Memory drop is the arena; handle-table drop is resource cleanup
  (this generalizes and replaces the per-connector
  `cleanup_*_connections()` pattern — RECOMMENDATIONS.md 1.7 / 5.1 become
  structurally unrepresentable).
- **One unit = one PIC wasm module.** Compile, cache, and invalidation
  granularity stay per-file. The `.uce → C++` translation pipeline is
  unchanged; only the compile target changes
  (`clang --target=wasm32-wasi -fPIC` + `wasm-ld -shared`).
- **Strictly lazy loading.** A unit's module is instantiated into a
  workspace the first time that workspace calls it — including mid-request.
  This is the wasm equivalent of today's compile-and-`dlopen`-on-first-hit
  and requires no restart, no fallback path. Placement memoization
  (deterministic bases so repeat instantiation is cheaper) is a permitted
  optimization; it must not change the loading policy.
- **In-flight isolation.** Module versions are immutable; a recompiled unit
  becomes a new module. Running workspaces keep what they loaded; new
  workspaces get the new version. Old modules are dropped when unreferenced
  (safe unload — impossible with `dlclose`).

### 3.2 Memory model

- **One heap, one allocator, one DValue implementation** — all owned by the
  core module. Unit modules *import* `malloc`/`free`/runtime symbols via GOT;
  the loader **rejects any unit module that defines rather than imports
  them** (two allocators on one heap is the one fatal misconfiguration).
- **Arena workspace allocator.** Because the workspace heap is dropped wholesale,
  the core's allocator may be a bump allocator with no-op free — the
  `_scratchpad.cpp` design, now correct by construction. Per-deployment
  flag; fallback is wasi-libc dlmalloc. Memory stats = heap pointer − base
  (replaces the tracking `operator new` in `types.h`).
- **Unit statics reset per request** (workspace is born from the core
  snapshot, which does not include unit data; unit data segments initialize
  at unit load within the workspace). This is *more* shared-nothing than
  today, where `.so` statics persist across requests within a worker.
  Cross-request state must use explicit host facilities (sessions, caches).
  **Breaking change — must be called out in docs and checked against the
  site/ tree during Phase 5.**

### 3.3 The host membrane

Exactly three currencies cross between host and workspace:

1. **scalars** (i32/i64/f64),
2. **byte buffers** (`ptr+len` into linear memory; inbound buffers are
   placed via the core's exported allocator), on the C++ side we strictly
   prefer the binary-safe std::string as a container for buffers
3. **handles** (opaque `u32` indices into the per-workspace host handle
   table; each entry carries a closer callback).

Everything pointer-shaped stays on its own side. Hostcall surface budget:
30–60 functions (§5.1). Host errors return as error values; traps are
reserved for unit faults. Nothing throws across the membrane.

**DValues cross the membrane only as the versioned wire encoding** (§5.3),
at the following sites: request context in (once), response out (once),
some hostcalls.

It is expected that many if not most of the toolset API functions we
expose to the unit developer will be judiciously split across the membrane:
have a relatively minimal wasm part that calls into the runtime host, and
then the runtime implementation which does most of the work.

### 3.4 DValue inside the workspace: no serialization, ever

Within a workspace, all modules share one address space, one toolchain, one
set of headers — the C/C++ ABI is intact across module boundaries. Therefore:

- A `DValue` is a pointer (an i32 offset into linear memory).
- `component(path, context)` resolves path → table index (host registry or
  guest-resident map) and `call_indirect`s, passing the context pointer.
  Reference semantics, mutation visibility, shared ob stack, working `ONCE`
  — identical to today.
- Function pointers are shared-table indices, valid across modules: virtual
  calls, `std::function` callbacks (`dv_map` lambdas) work across units.
- Cost: one `call_indirect` (single-digit ns) + GOT loads for cross-module
  symbols — the same shape of overhead native PIC pays through the PLT/GOT,
  i.e. what dlopened `.so` units pay today.

Encode/decode is **not** part of internal component calls. It exists only at
the membrane (§3.3) and on the cross-instance plane (§4).

### 3.5 The DValue C ABI (load-bearing, build it first)

The stable contract of the workspace is a **C ABI**, not the C++ class:
the core exports `extern "C"` accessors over an opaque `uce_dvalue*` (§5.2),
plus the string/ob/print helpers. C++ units may bypass it and use the class
directly (same headers, zero cost — a private fast path). Every other
workspace language uses the C surface.

This ABI is versioned: every unit artifact carries a custom section
(`uce.abi`: core ABI version + toolchain fingerprint); the loader refuses
stale units and triggers lazy recompilation (units are lazily compiled
anyway, so this costs nothing structurally).

**Phase 1 of the implementation plan is to introduce this C ABI in the
current native runtime** — it is useful immediately (plugin surface,
testability) and de-risks the rest.

---

## 4. Two component-call planes / multi-language support

The design stratifies languages by one question: *can the toolchain produce a
PIC linear-memory module that adopts a foreign allocator?*

**Plane A — workspace peers** (C++, C, Rust, Zig, …):
- Join the workspace as PIC modules importing core symbols.
- Must adopt the core allocator (Rust: `#[global_allocator]`; Zig: allocator
  parameter) and must not unwind across boundaries (`panic=abort` /
  catch-at-edge).
- Access DValues through the C ABI: pointer semantics, no copies, ns-scale
  calls. Idiomatic wrappers per language (e.g. Rust `DValue<'request>` —
  the borrow checker enforces the arena invariant).

**Plane B — runtime-carrying languages** (JS, Python, Go, C#, …):
- No support currently planned.

**The semantic rule (enforced by the loader, not by convention):**
cross-plane components do **not** receive the mutable context. They get an
explicit interface — props in (copied by definition), rendered output and
declared results back. Plane A keeps the full "here's the world, mutate it"
contract. A `component()` call must never silently change mutation semantics
based on the callee's implementation language.

The cross-instance call mechanism is shared by: Plane B units, and future
cross-trust-boundary components (multi-tenant). Serialization boundaries and
isolation boundaries are the same lines, by design.

---

## 5. ABI sketches (to be finalized in Phase 0/1)

### 5.1 Hostcall surface (grouped; target ≤ 60 functions)

```
request:   uce_host_ctx_read(buf) → len          // wire-encoded context, once
response:  uce_host_respond(status, hdrs_buf, body_buf)
           uce_host_stream_write(buf)            // chunked/streaming path
log:       uce_host_log(level, buf)
sqlite:    uce_host_sqlite_connect(path_buf) → handle | err
           uce_host_sqlite_query(handle, sql_buf, params_buf) → result_buf | err
           uce_host_sqlite_cursor_*(...)         // optional row-cursor variant
           uce_host_sqlite_insert_id/affected/error/disconnect(handle)
mysql:     (same shape; existing connector APIs are already handle-shaped)
files:     uce_host_file_read/write/stat/list(path_buf, ...)   // policy-gated
session:   uce_host_session_get/set(key_buf, val_buf)
http:      uce_host_http_request(req_buf) → handle/result_buf  // outbound
misc:      uce_host_time(), uce_host_random(buf), uce_host_env(key_buf)
loader:    uce_host_component_resolve(path_buf) → table_index  // may load (§6)
ws:        uce_host_ws_send(buf), event delivery via render entry re-invocation
```

Conventions: all errors as result codes + `uce_host_last_error(buf)`;
inbound buffers placed via the core's exported `uce_alloc`; no hostcall
traps on bad input (clamp/error instead).

### 5.2 DValue C ABI (core exports; sketch)

```c
typedef struct uce_dvalue uce_dvalue;                 // opaque; workspace-owned

uce_dvalue*  uce_dv_root(void);                   // request context
uce_dvalue*  uce_dv_get(uce_dvalue*, const char* key, size_t klen); // create-on-write
uce_dvalue*  uce_dv_find(uce_dvalue*, const char* key, size_t klen); // NULL if absent
const char* uce_dv_value(uce_dvalue*, size_t* len);
void        uce_dv_set_value(uce_dvalue*, const char* v, size_t vlen);
size_t      uce_dv_count(uce_dvalue*);
int         uce_dv_is_list(uce_dvalue*);
/* iteration */
uce_dv_iter uce_dv_iter_begin(uce_dvalue*);
int         uce_dv_iter_next(uce_dvalue*, uce_dv_iter*,
                                const char** key, size_t* klen, uce_dvalue** child);
/* encode/decode at the membrane */
size_t      uce_dv_encode(uce_dvalue*, char* buf, size_t cap);   // → UCEB1
uce_dvalue*  uce_dv_decode(const char* buf, size_t len);
/* ob / print / helpers: uce_print, uce_ob_start, uce_ob_get_close,
   uce_html_escape, uce_json_encode, ... (mirror uce_lib surface) */
```

No unwinding across this surface; C++ exceptions are caught at the edge and
surfaced as error returns where fallible.

### 5.3 Wire encoding "UCEB1" (membrane + cross-instance only)

Length-prefixed binary tree; **not** JSON. Sketch (finalize against DValue's
actual fields — value + ordered children):

```
node  := value children
value := varint len, bytes (utf-8)
children := varint count, count × ( key: varint len + bytes, node )
flags := one leading byte per node reserved (bit0: is_list hint)
header := "UCEB" u8 version
```

This encoding is a **versioned protocol** from day one (header byte). It is
the Plane B contract and the membrane format; internal calls never see it.

UCEB1 encoding/decoding should also be exposed to the unit developer so
they can make use of fast serialization/deserialization: matching our existing
API conventions these should be ucb_encode(DValue val) and ucb_decode(String val). This may also be
a worthwhile target for session variables storage (either change session
to DValue or add StringMap support to UCEB1 ser/de).

### 5.4 Unit module contract

```
custom sections:  dylink.0 (standard), uce.abi { abi_version, toolchain_id }
imports:          env.memory, env.__indirect_function_table,
                  env.__memory_base, env.__table_base,
                  GOT.mem.* / GOT.func.* (resolved by loader),
                  core symbols (malloc, uce_dv_*, uce_print, ...)
exports:          uce_unit_setup, uce_unit_render,
                  uce_unit_component, uce_unit_websocket
                  (same roles as today's UCE_SETUP/RENDER/COMPONENT/WEBSOCKET
                   dlsym symbols in compiler.cpp)
forbidden:        defining malloc/free/operator new, own memory, start fn
                  with side effects beyond data init
```

---

## 6. The loader (host-side, custom, load-bearing)

Owned code, ~1–2k lines, vendored-runtime-adjacent. Reference logic:
Emscripten's dylink loader (the ABI is the stable, battle-tested part; the
server-side loader is what doesn't exist off the shelf).

Per `load(unit)` into a workspace:

1. Fetch compiled module from artifact cache (compile on miss — today's
   lazy-compile path, retargeted).
2. Verify `uce.abi` stamp against the core; on mismatch, recompile unit.
3. Verify import discipline (no allocator/runtime definitions; §3.2).
4. Parse `dylink.0`: data size/alignment, table slots needed.
5. Allocate `__memory_base` (bump within workspace data region) and
   `__table_base` (append to shared table).
6. Instantiate with bases; resolve `GOT.*` imports against the workspace
   symbol registry (core symbols + previously loaded units); register the
   unit's exports. NB: data exports of PIC modules are `__memory_base`-relative
   offsets — add the owning unit's base when registering/resolving (core
   exports are absolute; the core is non-PIC). See the Phase 0 FINDINGS
   erratum; the Phase 3 spike's `self-got` marker exists to catch this.
7. Register entry points in the path → table-index dispatch map.

`uce_host_component_resolve(path)` consults the dispatch map and calls
`load()` on miss — this is how lazy, programmatic, mid-request loading works
with no special cases.

Placement memoization (optional, later): record each unit's first-assigned
bases; reuse across workspaces so instantiation is cheaper and snapshot
growth (below) stays consistent. Does not change the lazy policy.

**Core snapshot:** the only pre-built state is "core module, initialized" —
memory bytes + table state captured once per core build. Workspaces are born
from it via CoW (`mmap(MAP_PRIVATE)` of the snapshot image; the host owns
the Memory object, so OS-level CoW is available). No units are pre-fed.

---

## 7. Request lifecycle (replaces the native flow in linux_fastcgi.cpp)

```
1. accept request (fastcgi, websockets message, socket event, CLI request)
2. workspace = birth_from_core_snapshot()          (CoW, ~µs)
3. write wire-encoded context into workspace; core decodes → context DValue
4. resolve entry unit (load on first call); call uce_unit_render(ctx_ptr)
5. component(path) inside guest → resolve hostcall → (lazy load) →
   call_indirect — reference semantics throughout
6. I/O via hostcalls; resources land in the workspace handle table
7. on return: encode response/headers out; write FastCGI response
   on trap: defined error → render error page with guest stack trace;
   workspace state is irrelevant because…
8. drop workspace: linear memory gone (arena), handle table closed
   (generalized resource cleanup), instances released
```

CPU limit: epoch/fuel interruption → same path as trap.
Memory limit: linear memory max → allocation failure / trap → same path.

---

## 8. What carries over unchanged

- The `.uce → C++` translation, parser, and page semantics.
- The lazy compile-on-first-request model and per-unit artifact caching
  (different artifact format).
- The host-side connectors (sqlite/mysql) — already handle-shaped APIs; they
  move behind hostcalls with the same `.uce`-visible signatures.
- The site tree, docs, demo, and the network test suite
  (`tests/run_network_tests.py`) — which becomes the parity harness (§9).
- nginx/FastCGI front-end integration, worker model, websocket event flow
  (events re-enter via the websocket entry point).

---

## 9. Implementation plan

Phases are sequential; each has an exit criterion. No phase except 0 and 2's
scaffolding produces throwaway work. All dependencies must be vendored.

**Phase 0 — toolchain & runtime.**
Validate: wasi-sdk `-fPIC` + `wasm-ld -shared` on a representative generated
unit; cross-module C++ calls with shared memory/table; exceptions decision
(wasm EH vs. error-code discipline at unit boundaries — pick one, record it);
vendored runtime selection. Candidates: **WAMR** (C, small, designed for
embedding, easiest to vendor and patch — fits the project's vendoring
practice and would be preferred) vs. **Wasmtime** (fastest, best AOT/CoW machinery, Rust — heavier
to vendor/patch, use only if blocked on WAMR). Selection criteria: imported-memory + shared-table support,
AOT artifact quality, patchability. Exit: a two-module (core stub + unit
stub) hello-world linked at runtime by a minimal loader, in the chosen
vendored runtime.

> **Status: DONE (2026-06-12).** Exit criterion passed on k-uce; see
> `spikes/wasm-phase0/FINDINGS.md`. Runtime selected: **Wasmtime v45.0.1**
> — WAMR is blocked on the load-bearing requirement (its wasm-c-api ignores
> imported memories/tables; host-side table growth unsupported). wasi-sdk-33
> PIC validated on stubs **and** on real generated units
> (`collections.uce.cpp`, `hello.uce.cpp` → side modules, no allocator
> definitions). Exceptions: `-fno-exceptions` confirmed across all of it
> (§11.1 stands). Cross-module C++ (containers, heap objects, function
> pointers, `std::function` lambdas, GOT.mem/GOT.func) all proven through
> the spike loader.

**Phase 1 — DValue C ABI in the native runtime.**
Introduce `uce_dv_*` (§5.2) and the UCEB1 codec in `src/lib/`, used
natively. Zero wasm dependency; immediately testable; freezes the contract
everything else builds on. Exit: codec round-trip + accessor tests in the
existing suite; ABI doc checked in.

> **Status: DONE (2026-06-12).** Native runtime now exposes `uce_dv_*`
> accessors and UCEB1 encode/decode helpers in `src/lib/dvalue.{h,cpp}`.
> ABI details are checked in at `docs/wasm-phase1-dvalue-abi.md`; UCE-visible
> docs are available as `ucb_encode`/`ucb_decode`. Exit coverage is in
> `site/tests/core.uce` and passed in the full network suite.

***Phase 1 Addendum***

  Fix before commit:

1. ucb_encode(DValue value) deep-copies the whole tree (dvalue.cpp:962, same signature in the header). DValue copy is a full recursive map+string clone, and this function is the future membrane hot path — the request context will pass through it on every request in Phase 2. Should be const DValue& (the function only reads). Same nit for bool ucb_decode(String encoded, ...) at :971 — a by-value String copy of what may
be a large document; const String& matches.

2. 'P' values ship the raw pointer address on the wire (ucb_node_scalar, dvalue.cpp:833, the 'P' case). The ABI doc explicitly says "pointer/reference identity is intentionally not part of the wire contract," but the implementation encodes std::to_string((u64)ptr) — a meaningless number on the receiving side and an ASLR address disclosure the day UCEB1 crosses a trust boundary (Plane B / multi-tenant is the stated
endgame). It's consistent with native to_string, but the wire is a different context: I'd encode "" for 'P' and note it in the doc.

3. f64 fidelity on the wire. 'F' encodes through std::to_string → fixed 6 decimals. That's faithful to native to_string, but the membrane makes it new lossiness: today an 'F' value never round-trips through its string form unless page code asks; in Phase 2 every float in the context will. 1e-7 becomes "0.000000" → decodes to 0. Since the scalar is just a string, switching 'F' to shortest-round-trip formatting
(%.17g-style) later needs no version bump.

4. uce_dv_iter is about to be frozen with no headroom. Keyed-map iteration does std::advance(begin(), position) per call (dvalue.cpp:1096) — O(n²) per full sweep, and the C ABI will be the only iteration path for non-C++ units. The fix (e.g. resuming via lower_bound on the last key) needs state the one-field struct can't hold. Phase 1's whole purpose is freezing this contract: I'd add reserved space now (size_t
position; size_t reserved[3]; or an opaque byte array) so the implementation can get smarter without an ABI break.

Also:

  - uce_dv_decode returns a pointer into a single thread-local slot (:1122) — a second decode silently invalidates the first result. The doc documents borrowing for uce_dv_value but not this; one sentence ("valid until the next uce_dv_decode on the thread") would close it.
  - An empty non-list map round-trips as scalar "" (type 'M' → 'S'; the child_count == 0 && !LIST branch at dvalue.cpp:873ff). HOPEFULLY harmless in practice (is_array() flips), maybe worth a doc line.
  - The decoder silently drops a scalar when children are present — unreachable from the encoder, only crafted input. Fine for v1; "reserved" mention in the format doc would pin it.
  - Tests cover only the happy path. The hardening (truncation, bad magic, wrong version, depth bomb) is implemented but untested — two or three negative ucb_decode checks in core.uce would lock it in. A float/bool round-trip check would also have surfaced finding 3.

> **Addendum status: DONE (2026-06-12).** `ucb_encode`/`ucb_decode` now take
> const references, pointer nodes encode as empty scalars, floating-point
> scalars use `max_digits10`, `uce_dv_iter` has reserved ABI headroom, docs
> cover decode-root lifetime and v1 edge cases, and core tests include invalid
> input plus float/bool round-trips.

**Phase 2 — core module + membrane.**
Compile `uce_lib` (+ wasi-libc) to wasm as the core module; implement the
hostcall surface (§5.1) in the host; temporary scaffolding allowed: one
statically-linked unit + core to validate codegen and membrane without the
loader. Exit: one real `.uce` page (e.g. `site/tests/core.uce`) renders
correctly through the membrane. Scaffolding is marked throwaway.

> **Status: MEMBRANE SCAFFOLD DONE (2026-06-12).** Temporary scaffold checked
> in under `spikes/wasm-phase2/`: a WASM reactor core subset owns memory,
> `Request`, `DValue`, UCEB1, and output buffering; the Wasmtime host implements
> the initial `uce_host_ctx_read`/`uce_host_log` membrane and passes a UCEB1
> request context into the guest; a statically linked `.uce` render entry runs
> through that membrane. Validation passed on k-uce with `PHASE2 EXIT CRITERION:
> PASS`. The scaffold does not yet exercise UCE preprocessor-emitted page C++;
> generator-emitted units through this membrane are explicitly deferred into the
> Phase 3 loader path. Full dynamic unit loading also remains Phase 3 as planned.

**Phase 3 — the loader + workspace.**
Implement §6 in full: dylink parsing, base allocation, GOT resolution,
ABI/import verification, lazy mid-request loading, path dispatch. Per-request
workspace birth/drop (plain memcpy birth is fine here; CoW is Phase 4).
Exit: the uce-starter renders end-to-end with components loading lazily;
`tests/run_network_tests.py --match starter` passes against the wasm worker.

> **Status: SPIKE PASS (2026-06-12).** `spikes/wasm-phase3/` combines the
> Phase 0 dylink/PIC loader with the Phase 2 UCEB1 context membrane. The spike
> builds a core workspace reactor and a separate generated-shape `.uce.cpp` PIC
> side module, parses `dylink.0`, allocates memory/table bases, resolves
> `env.*`/`GOT.*` imports, runs relocations/constructors, calls
> `__uce_set_current_request` and `__uce_render`, and reads output from
> core-owned memory. The fixture now exercises nonzero table placement,
> `GOT.func`, deferred self-resolved `GOT.mem`, and generated-style
> `html_escape(...)` expression output. It passed on k-uce with `PHASE3 EXIT
> CRITERION: PASS`.
>
> The self-GOT fixture caught a real loader bug (2026-06-12, fixed): deferred
> `GOT.mem` entries were patched with the unit's exported symbol values
> verbatim, but a PIC module's data exports are offsets relative to its
> `__memory_base` — the loader must add the base. The bug rendered silently
> wrong values and wrote into core memory at low addresses; both spike loaders
> are fixed and the Phase 3 exit gate now asserts every GOT-derived output
> marker (`self-got`/`callback`/`each`/`map`) so a regression cannot pass.
> Details in the `spikes/wasm-phase0/FINDINGS.md` erratum.
>
> **Production Phase 3 work plan** (ordered by dependency/risk; spike-proven
> mechanics not repeated here):
>
> 1. **Compile the real `uce_lib` as `core.wasm`** — the last big unknown; all
>    spikes used a hand-stubbed `Request`. Forces: `types.h` allocator gate as
>    a real `#ifdef` (replacing the spike's copied-header text patch), `sys.h`
>    signal/fork/socket carve-outs, the libc++ closure strategy
>    (`--whole-archive` vs keep-list), and splitting connectors out of the core
>    (the MySQL client library cannot compile to wasm). Gates items 2–6.
> 2. **WASI decision (record in §11 when made)** — spikes stub all WASI imports
>    with traps; real pages call `time()` (7× in uce-starter), which wasi-libc
>    routes to `clock_time_get`. Recommended: zero-WASI core — route
>    time/random/env through the §5.1 hostcalls so there is exactly one
>    membrane to audit.
> 3. **Generator changes** (small, parallelizable): emit a logical include
>    instead of the absolute `uce_lib.h` path; add the PIC side-module build to
>    the compile-on-miss artifact cache.
> 4. **Productionize the loader into `src/`** (§6): `uce.abi` stamping,
>    import-discipline verification, a name→funcptr registry in the core
>    replacing the spike's per-symbol `core_table_index_of_*` helpers, an
>    explicit export name-collision policy (the spike silently prefers core
>    exports over unit definitions, e.g. `context`), multi-unit bump placement,
>    retained unaligned allocation pointers for unload, hardened/fuzzed binary
>    parsing. Plain memcpy workspace birth; CoW stays Phase 4.
> 5. **Starter-scoped hostcalls (~12, not the full ≤60)**: `respond` /
>    `stream_write`, `time`/`random`/`env`, `session_get`/`set`,
>    `http_request` (OAuth callback), `component_resolve`. The starter uses no
>    sqlite/mysql/file APIs — connectors can wait for Phase 5 parity.
> 6. **`component_resolve` + path dispatch + lazy loading** — the only §6 step
>    with zero spike coverage (all spikes load one unit eagerly). The starter's
>    71 `component()` calls across 47 files are the stress test; worth a
>    focused spike before worker integration.
> 7. **FastCGI worker integration** — config-selectable backend, §7 lifecycle
>    wired into the `linux_fastcgi.cpp` flow.
> 8. **Starter parity tests — DONE (2026-06-12).**
>    `tests/plugins/uce_starter_parity.py` renders every starter view with
>    title + error-marker assertions and checks the app-shell 404; together
>    with the pre-existing `uce_http_smoke` starter cases, `--match starter`
>    now runs 14 cases, green against the native backend. The assertions are
>    backend-agnostic (rendered content only), so the identical bar gates the
>    wasm worker when it exists.

**Phase 4 — production mechanics.**
Core snapshot + CoW birth; bump-allocator flag; epoch/memory limits; trap →
error-page path with guest stack traces (this supersedes the
signal/longjmp machinery and closes RECOMMENDATIONS.md 1.5 structurally);
handle-table cleanup (closes 1.7/5.1); artifact/ABI versioning end-to-end.
Exit: kill-tests (deliberate out-of-bounds page, stack-exhaustion page,
infinite-loop page, OOM page) produce clean error pages and an unharmed
worker. (A literal null-deref page is not in the list: address 0 is valid
wasm linear memory, so it does not trap — see §10.)

> **Status: KILL-TEST SPIKE PASS (2026-06-12).** `spikes/wasm-phase4/`
> validates the production mechanics that can be proven before the real wasm
> worker exists: reusable compiled artifacts as a core-snapshot proxy, fresh
> per-request stores as workspace birth/drop, **both** CPU-limit mechanisms
> (fuel and epoch interruption — production default is epoch per the Phase 0
> findings; Wasmtime reports epoch traps as `interrupt`), a store memory
> limiter proven load-bearing (the OOM fixture grows within its own declared
> max so only the limiter can deny it), trap capture with the exit gate
> asserting a wasm backtrace and the expected cause per kill, handle closers
> checked to run exactly once and while the store is alive, and — after all
> six kills — a healthy request served by the same engine (the "unharmed
> worker" half of the exit criterion). Kill fixtures are genuinely trapping
> faults: unreachable, OOB access, stack exhaustion, fuel/epoch-limited
> infinite loops, limiter-denied growth. A literal null deref is deliberately
> absent (does not trap in wasm; risk recorded in §10). Passed on k-uce with
> `PHASE4 EXIT CRITERION: PASS`. The trap-trace summarizer now exists as
> `src/lib/wasm_trace.h` (collapses repeated frames, demangles symbols,
> splits cause/detail) and is double-gated: live traps in the spike runner,
> canned-message checks in `site/tests/core.uce` via the native suite. Not
> yet production: OS-level CoW core snapshots, wiring trace summaries into
> the error-page UI, the unit name-section policy for readable frames,
> `linux_fastcgi.cpp` wiring, real connector handle cleanup, and artifact/ABI
> versioning.

**Phase 5 — parity & performance.**
Full network suite green on the wasm worker; differential native-vs-wasm runs
on the site tree; audit `site/` for cross-request-static reliance (§3.2
breaking change); benchmark suite (template-heavy page, sqlite page,
component-heavy starter page) with budgets: ≤2× native page latency,
workspace birth ≤100µs, internal component call overhead within 10× native
call cost. Exit: numbers published in this document, all tests and reviews pass.

> **Status: HARNESS BASELINE PASS (2026-06-12).** `spikes/wasm-phase5/`
> now automates the Phase 5 parity/performance gate shape before the production
> wasm worker exists. On k-uce it ran the full native network suite (`83/83`),
> the starter-focused parity subset (`14/14`), a heuristic code-focused `site/`
> cross-request/static-state audit, and a warmed native benchmark baseline for
> the template-heavy doc page, sqlite page, and component-heavy starter page.
> The harness now gates case counts (`network >= 80`, `starter >= 10`) so broken
> filters cannot pass vacuously, and it runs a throwaway warmup suite before the
> measured gate to absorb cold unit-cache timeouts. Informational native medians
> from the 2026-06-12 baseline are: template-heavy doc `313.8 ms`, sqlite page
> `3.4 ms`, starter dashboard `41.1 ms`. A durable snapshot lives in
> `spikes/wasm-phase5/reports/native-baseline-2026-06-12.md`; paired wasm/native
> gate runs still recompute the native medians for the actual ≤2× comparison.
> True Phase 5 completion still requires passing the same harness against a real
> wasm worker URL and adding worker-internal probes for workspace birth and
> component-call overhead budgets.

**Phase 6 — second plane (deferred until wanted).**
Cross-instance call mechanism (props-in/output-out, UCEB1), first Plane B
language binding, loader enforcement of the cross-plane context rule (§4).

The native `.so` backend remains in-tree (as a reference) and selectable by config
but we switch over to the wasm backend as soon as it's available and test
only on that; both backends share the Phase 1 C ABI.

---

## 10. Risks & mitigations

| Risk | Mitigation |
|---|---|
| wasi-sdk PIC / shared-library maturity (least-trodden toolchain path) | Phase 0 spike before any commitment; pin toolchain versions; statically link libc into the core and export from there (avoid shared wasi-libc entirely) |
| C++ exceptions × PIC × wasm EH | Phase 0 decision point; fallback is error-code discipline at unit entry points (units already have a uniform entry shape) |
| Custom loader correctness (GOT, bases, relocation) | Small, contained (~1–2k lines); crib logic from Emscripten's reference loader; fuzz with adversarial modules; loader rejects > loader guesses |
| Vendored runtime patches drift from upstream | Same practice as vendored SQLite: provenance + patch files under `docs/patches/`; pin upstream tag; tests gate upgrades |
| Performance regression beyond budget | Phase 5 gates; bump allocator and placement memoization in reserve; native backend retained |
| Multi-module DWARF / debugging story | Trap stack traces cover the production case (better than today); accept weaker interactive debugging; keep native backend for local deep-debugging |
| Unit-statics semantic change breaks existing pages | Phase 5 audit of `site/`; documented migration note; host-side cache facility if a real need surfaces |
| Null-pointer dereferences do not trap (wasm address 0 is valid linear memory) — a buggy page writes low workspace memory and renders silently wrong output instead of erroring | Accepted: damage is confined to the request's workspace and dropped at request end (strictly better than native SIGSEGV). Kill-tests use genuinely trapping faults (OOB, stack exhaustion, `__builtin_trap`). Optional later: null-check instrumentation at a measured perf cost |

## 11. Decisions

1. **Exceptions:** wasm EH or error codes at unit boundaries? Error codes.
2. **WebSocket granularity:** workspace per event (pure arena, statics reset
   per event) or per connection (state across events, bounded lifetime)?
   Per-event. Should be compatible with WS since the WS contract is: runtime holds
   and brokers connections, makes request to unit's WS() {...} directive, unit
   may decide to send data back over WS to any or even all connected clients.
3. **UCEB1 final layout** vs. DValue's actual field set (value/children/list
   flag) — somewhat open, finalize in Phase 1.
4. **Cursor vs. bulk** as the default for `sqlite_query` results at the
   membrane (offer both; pick the default after Phase 5 benchmarks, tending towards cursor right now).
5. **Streaming output:** today's ob model buffers; does the membrane expose
   `uce_host_stream_write` from day one or post-MVP? Undecided. ob is a critical
   abstraction to our whole execution model and whether we expose direct stream
   write or not, the existing PHP-like ob contract/paradigm must stay in place.
6. **Core snapshot rebuild cadence** once placement memoization lands
   (dead-base reclamation policy).

## 12. Summary

The module is the **unit** (file-grained, lazily compiled, lazily loaded —
including mid-request). The instance set is the **workspace** (one per
request; shared memory + table; born from a core-only CoW snapshot; dropped
wholesale — the arena, done right). The contract is the **DValue C ABI**
inside the workspace (pointer semantics, no serialization) and the **UCEB1
wire encoding + handles** at every true address-space boundary (host
membrane, Plane B languages, future trust boundaries). Serialization
boundaries and isolation boundaries are the same lines; component calls stay
function calls; and the file stays the unit.