rose-ash/plans/sx-review/conformance.md

# SX Conformance Review — cross-host agreement + test adequacy

Axis: CONFORMANCE (do the hosts agree, and would the suite catch it if not?). Sibling lanes: core semantics, host implementations.
Date: 2026-07-03. All test runs and probes executed this session (bounded with `timeout`; no shared sx_server touched — probes used freshly-spawned bounded `sx_server.exe` instances).
CONFIRMED = reproduced here; SUSPECTED = static reasoning, probe proposed. Severity S1 critical → S4 minor.
Probe artifacts: `/tmp/claude-0/-root-rose-ash/9a04ba52-7bf4-476d-99ea-04f84bff1359/scratchpad/{probes,prims}/`; raw suite logs `/tmp/sx-review/*.log`.

---

## 0. Ground truth: what the hosts actually are (the briefed picture is stale)

- **OCaml native kernel** (`hosts/ocaml`, sx_server.exe) — canonical evaluator, essentially green on the corpus.
- **OCaml WASM kernel** (`shared/static/wasm/sx_browser.bc.wasm.js`, js_of_ocaml build of the same kernel) — **what production browsers actually run** (served at lib/host/blog.sx:1910). Served artifact verified identical (md5) to freshly built one — no artifact staleness.
- **JS-transpiled bundle** (`hosts/javascript` → `shared/static/scripts/sx-browser.js`) — legacy bootstrapped evaluator; still built and gated by `scripts/sx-build-all.sh`; catastrophically red (below). Not referenced by any served page found.
- **Python host** (`hosts/python`) — vestigial: test runner deleted in d735e28b; `SX_USE_OCAML=1` in every compose file; bootstrap output consumed by nothing. NOT a live evaluator target. BUT `shared/sx/parser.py` (an independent Python SX reader/serializer) **is live in production plumbing** — see F-8.
- **hosts/native** — Cairo pixel-renderer (separate host concept), single smoke test, not a corpus runner.

## Suite results (this session)

| Runner | Briefed | Actual |
|---|---|---|
| JS standard (`node hosts/javascript/run_tests.js`) | ~747 green | **2596 pass / 2490 FAIL** (5086) |
| JS full (`--full`) | ~870 green | **2453 pass / 3203 FAIL** (5656) |
| Python | ~744 green | **runner deleted** (d735e28b) — zero tests |
| OCaml (`run_tests.exe`; no `--full` flag exists) | ~1080 green | **5762 pass / 274 FAIL** |
| OCaml WASM kernel | — | **corpus never runs on it** (F-2) |

Rebuilding the JS bundle fresh from today's spec reproduces the JS failures exactly (2596/2490) → red JS is real, not a stale artifact.

---

# CONFIRMED findings (most severe first)

## F-1 [S1, confidence high] Server and browser disagree on integer arithmetic — silently
The WASM kernel (the artifact every production browser loads) has 32-bit int semantics (js_of_ocaml); native is 63-bit. Same expression, same "canonical kernel", different answers, no error:

| Expr | native sx_server | WASM SxKernel (shipped) | legacy JS bundle |
|---|---|---|---|
| `(* 99999999 99999999)` | 9999999800000001 | **1674919425** (32-bit wrap) | 9999999800000000 (float64) |
| `(+ 9007199254740992 1)` | 9007199254740993 | **0** (literal truncated mod 2^32) | 9007199254740992 |
| `(expt 2 62)` | -4611686018427387904 (int63 wrap) | **0** | 4611686018427388000 (float) |
| `(* 999999999999 999999999999)` | 2003762205206896641 (silent int63 wrap) | 9.99999999998e+23 (goes float) | 9.99999999998e+23 |

Three hosts, three different answers on the same input. Repro: `scratchpad/probes/run_wasm.js probes.txt` vs `sx_server.exe (eval ...)`. Any SSR-computed value re-derived client-side (pagination totals, ids, hashes-by-arithmetic, seat counts) can differ. Note native has two hazards of its own: silent int63 wraparound and reduced float print precision (`(+ 0.1 0.2)` prints `0.3`).
Fix ownership: hosts lane (int64/boxed ints or explicit overflow policy in the WASM build); THIS lane's ask: a numeric-tower differential suite that runs on every shipped artifact.

## F-2 [S1, high] The shipped browser kernel never runs the corpus
No runner feeds spec/tests through `sx_browser.bc.js`/`.wasm.js`. Its entire test surface: `test_boot.sh` (require() smoke), `test_wasm_native.js` (test-framework.sx + web/tests/test-wasm-browser.sx only), `tests/node/run-sx-tests.js` (one deftest file). Both "conformance" runners test kernels that are NOT the shipped browser artifact. F-1 and F-3 existed undetected precisely because of this. The probe harness at `scratchpad/probes/run_wasm.js` shows a corpus runner on the WASM kernel is trivially buildable (SxKernel.eval works headless in node with the test_boot.sh stub block).

## F-3 [S1, high] Native and WASM disagree on `apply` and dict key order (same kernel family!)
- `(apply + (list 1 2 3))` → native: **error** "Expected number, got list" (apply does not spread); WASM: **6** (spreads, 2-arg form only; `(apply + 1 (list 2 3))` errors "apply: function and list"). Legacy JS bundle spreads fully (both forms → 6). Three different apply semantics.
- `(assoc {:a 1} :b 2)` → native `{:b 2 :a 1}` (new keys prepend); WASM `{:a 1 :b 2}` (insertion order); JS bundle insertion order. `merge` same. Dict iteration/serialization order differs server-vs-browser — anything ordering-dependent (rendered attr order, keys/vals iteration, golden-file comparisons) diverges.
- Defused non-finding: `cid-from-sx` canonicalizes — CIDs of `{:a 1 :b 2}` and `(assoc {:a 1} :b 2)` are identical on native AND WASM (probe verified). Content addressing is order-safe.

## F-4 [S1, high] JS host fails ~half the shared corpus; the documented build gate is red
spec/tests discovery is byte-identical in both runners (82 files both), so results are directly comparable: OCaml ≈ green, JS 2490 FAIL / 5086. `scripts/sx-build-all.sh` (`set -euo pipefail`) runs `node hosts/javascript/run_tests.js --full` as a gate → **the documented full pipeline currently exits FAIL** (run_tests.js exits 1). Either the JS host is dead (remove from gate/docs) or alive (≈2500 tests behind). Today no automated gate enforces cross-host agreement at all.

## F-5 [S1, high] The corpus is not host-neutral: OCaml runner preloads libraries + services `import`; JS runner does neither
- OCaml `make_test_env` (run_tests.ml:3700-3806) unconditionally preloads r7rs/render/canonical/adapters, forms/engine/router/orchestration, bytecode/compiler/vm, stdlib, signals, freeze, content, **parser-combinators, graphql, graphql-exec**, dom/browser, and the full **hyperscript stack**.
- OCaml services `(import ...)` IO-suspensions at runtime (run_tests.ml:2148-2169, 2178-2272). The JS runner calls `Sx.eval` directly — **no suspension/resume loop**, so `import` can never complete on JS.
- Result: hundreds of "Undefined symbol: gql-tokenize / pc-* / hs-*" JS failures (js-standard.log) that are runner-environment artifacts layered on top of real gaps. Tests exercising import/suspension (test-import-bind, test-io-suspension, test-coroutines, test-modules) do not test the same thing per host.

## F-6 [S1, high] Whole test directories are gated on only one host — or none
From run_tests.js:300-448 and run_tests.ml:3959-4001:
- **lib/tests (11 files)** — continuations, continuations-advanced, freeze, signals-advanced, stdlib, stepper, tree-tools, types, vm, vm-closures, vm-primitives — auto-run **only on JS `--full`** (OCaml runs them only if named explicitly). Their only automated gate is the runner that's red: e.g. test-continuations.sx on JS: 475 pass / 54 FAIL; algebraic-data-types: 53 FAIL. **Effectively ungated.**
- **web/tests (13 runnable)** — adapter-html, aser, deps, engine, examples, forms, orchestration, page-helpers, relate-picker, router, signals, swap-integration, tw-layout — auto-run **only on OCaml**.
- **6 web/tests run on NO standard runner**: test-adapter-dom, test-boot-helpers, test-cek-reactive, test-handlers (module-loaded, not run as suite), test-layout, test-wasm-browser (only via test_wasm_native.js).
- **OCaml foundation tests** (run_tests.ml:1178+) — native-only unit tests, no cross-host equivalent (fine, but they're invisible to other hosts by design).
- `spec/tests/test.sx` matches neither filter (`test-*.sx`) → **runs nowhere**.
- Guest-language suites (lib/scheme/tests/records.sx, lib/haskell/tests/records.sx, …) sit outside all discovery — each loop self-tests ad hoc, no aggregate gate.

## F-7 [S1, high] Features pass their tests only inside the test runner's private environment — both directions
- **OCaml side**: `values`, `promise?`, `make-promise`, `force` are bound **only in run_tests.ml** (lines 1131-1167). Probes: `(force (delay (+ 1 2)))`, `(values ...)`, `(let-values ...)` → "Undefined symbol" on BOTH production surfaces (native sx_server epoch env AND WASM SxKernel). ~271 promises/values assertions pass against an environment that exists nowhere in production.
- **JS side**: run_tests.js injects `equal?`, `apply`, `env-*`, `render-html`, `make-continuation`, upcase/downcase, and a **fake sha3-256 stub** (lines 80-160) into the test env. The real browser bundle lacks these (probe: `equal?` → Undefined symbol on the bundle). Coverage numbers overstate both shipped artifacts.

## F-8 [S1→S3 itemized, high] Differential probes, OCaml native vs JS bundle: 130 exprs, 98 identical, real divergences
(harness: `scratchpad/probes/`; numeric rows already in F-1)

Core data semantics:
| Expr | OCaml native | JS bundle |
|---|---|---|
| `(cons 1 nil)` | `(1)` (nil ≡ empty list) | `(1 nil)` (nil is an element) — **foundational** |
| `(range 3)` | `(0 1 2)` | `()` — single-arg range returns empty on JS |
| `(round -2.5)` | -3 (half away from zero) | -2 (JS half-up) |
| `(str (list 1 2 3))` | `"(1 2 3)"` | `"1,2,3"` (Array.toString leak) |
| `(str {:a 1})` | `"{:a 1}"` | `"[object Object]"` |
| `(len "héllo")` / `(len "👍")` | 6 / 4 (bytes) | 5 / 2 (UTF-16 units) — string indexing math differs on all non-ASCII; neither counts codepoints |
| `(upper "straße")` | "STRAßE" (ASCII-only) | "STRASSE" (Unicode) |
| `(reverse "abc")` | error (lists only) | `"cba"` |
| `(sort lst cmp-fn)` | error "sort: 1 list" | sorts with comparator |
| `(max)` | error | `-Infinity` |
| `(/ 1 0)` | `inf` | `Infinity` (print) |
| `(+ 0.1 0.2)` | prints `0.3` | prints `0.30000000000000004` |

Agreements worth recording (hosts agree, docs/spec don't): `case` uses flat pairs `(case x 1 "one" 2 "two" :else ...)` — the CLAUDE.md-documented clause-list syntax errors identically on all three kernels; `(join sep coll)` is sep-first on both; `int?` undefined on both (it's `integer?`); `string->number` radix arg unsupported on all hosts (and the corpus asserts it — a live OCaml FAIL, see F-10).

## F-9 [S2, high] Primitive-set parity is broken at scale (full lists in `scratchpad/prims/report.txt`)
- **194 OCaml-core names missing from the JS bundle**, incl. language-level: `take drop zip unique init char-at substr upcase downcase equal? identical? dict-get dict-has? dict-delete! parse parse-safe parse-float escape-string compile compile-module` + records constructors + math extras. Dynamically verified samples throw "Undefined symbol" on JS.
- **Naming splits where both have the capability**: OCaml `regex-*` vs JS `regexp-*`; OCaml `parse` vs JS `sx-parse`; OCaml `json-encode` vs JS `json-stringify`.
- **24 user-facing JS names missing on the OCaml kernel**: `format values force promise? make-promise call-with-values json-parse json-stringify format-date parse-datetime pluralize escape strip-tags error-message env-parent char-code-at now-ms satisfies? …` (deps-check verified). Note the epoch-mode server doesn't load `spec/stdlib.sx`, so even `format` is unresolved in production server env.
- **Declared in spec/primitives.sx but implemented NOWHERE: `eq?`, `eqv?`** (202 spec-declared total; 9 missing from OCaml, 6 from JS).
- JS bundle oddities: 6 statically-assigned PRIMITIVES keys absent at runtime (`loop raw-loop reactive-text read-char-name-loop read-map-loop scan`).

## F-10 [S2, high] The OCaml suite is not green, and a permanent red band is normalized
274 FAIL on the canonical host: 272 hs-upstream-* (fetch/socket/runtimeErrors/asExpression/...). If browser-only, they should be skip-listed like the 6 web tests — a permanently red FAIL column trains everyone to ignore failures. The 2 core failures are live: `can-map-an-array` ("map with block") and `string->number` radix (corpus asserts a feature no host implements). Also note hs-upstream pass/fail sets differ wildly between OCaml (272 fail) and JS (~1900 fail) — the suites' shared-corpus value is currently nil for hyperscript. **UPDATE (S-2 probe, F-18): ≥118 of the 272 red hs tests PASS on the shipped WASM kernel + happy-dom — the red band is mostly mock-DOM environment deficiency, not engine failure.**

## F-11 [S2, high] Boundary validation silently disabled in production; stale imports broke the Python-side test files
- `shared/sx/boundary.py:34` imports `.ref.boundary_parser` — moved to hosts/python in 7036621b → ImportError swallowed (boundary.py:44-49) → **boundary type validation is a silent no-op**.
- `shared/sx/tests/test_bootstrapper.py:149`, `test_parity.py:733` import deleted `shared.sx.ref.*` — those tests cannot even load.
- `shared/sx/async_eval.py` (fallback evaluator) is macro-crippled by design (raises "sx_ref.py has been removed"); `resolver.py` fully stubbed. Fallback path SX_USE_OCAML=0 is non-functional — fine if intentional, but it's still importable wiring.
(Hand-off: hosts lane for the fix; kept here because it *is* the boundary-conformance enforcement.)

## F-12 [S2, high] The live Python SX reader diverges from the OCaml reader
`shared/sx/parser.py` runs on every production request (serialize/parse plumbing under SX_USE_OCAML=1: handlers.py:191,211; helpers.py:449). Probes vs OCaml reader:
- Python CANNOT read: `#t`/`#f` (reader-macro error), `#\a` chars, dotted pairs `(a . b)`, `.5` — all valid on OCaml. If any OCaml-serialized value containing a char (`#\a`) crosses the Python plumbing, Python errors.
- Dict serialization order differs (Python insertion `{:a 1 :b "two"}` vs OCaml reversed `{:b "two" :a 1}`) — hazard for any wire-text comparison/caching (CIDs themselves are safe, F-3).
- Agreements: escapes (\n \t \" \\), floats/exponents, keywords, `[..]`-as-list, quote/quasiquote sugar.

## F-13 [S3, med] The checked-in generated kernel is not reproducible from today's spec+bootstrap
`python3 hosts/ocaml/bootstrap.py --output <scratch>` vs checked-in `hosts/ocaml/lib/sx_ref.ml`: 360 diff lines. Function inventory is identical (the delta is `let` vs `and` restructuring + block moves — an older generator produced the checked-in file); no semantic difference detected, and tree==\_build. But CI (.gitea/Dockerfile.test steps 3-4) re-bootstraps sx_ref.ml from spec and recompiles — so CI's binary is built from different generated source than the dev tree's. Reproducibility should be a checked invariant (regen + diff in CI).

## F-14 [S3, high] Doc drift — CLAUDE.md describes a system that no longer exists
- ~~"Canonical SX semantics in `shared/sx/ref/*.sx`"~~ — **FIXED 2026-07-03**: CLAUDE.md now points at `spec/` and documents the bootstrap chain + corrected island rules (sequential `let`, implicit begin). Remaining items below still stand.
- Documents a Python host + ~744 tests — deleted.
- Documents `case` clause-list syntax — all hosts reject it (flat-pair syntax is real, spec/tests/test-cek-advanced.sx:549).
- Briefed suite counts (747/870/744/1080) are months stale; corpus is 5-6k assertions.
- `spec/primitives.sx` header claims stdlib functions "moved to stdlib" but spec/stdlib.sx defines only `format` (and production doesn't load it).

## F-15 [S4, med] Housekeeping
- Stray recursive tree `hosts/ocaml/hosts/ocaml/hosts/ocaml/bin/sx_server.ml` (accidental copy) — confuses greps and could get compiled/edited by mistake.
- `spec/tests/test.sx` dead filename (see F-6).
- JS runner's fake `sha3-256` stub returns non-SHA3 values — any hash-shape-only assertion passes; any real-value assertion would mysteriously fail JS-only.

---

# Hyperscript on the shipped WASM kernel (S-2 probe, executed 2026-07-03)

Method: new probe harness `scratchpad/hsprobe/run_hs_wasm.js` — loads the SHIPPED kernel (`shared/static/wasm/sx_browser.bc.js`) + shipped platform in node/happy-dom via `tests/node/sx-harness.js`, preloads the same files as the OCaml runner (test-framework, spec/harness, web/lib/dom, the lib/hyperscript stack), and runs the hs spec corpus with streamed per-test output. Per-test diff vs the native run in `scratchpad/hsprobe/compare2.py`; raw logs `/tmp/sx-review/wasm-hs-*.log`, comparisons `/tmp/sx-review/wasm-hs-compare.txt`, `wasm-hs-pure2-compare.txt`. Coverage: 1,553 tests measured (tokenizer/parser/compiler/runtime 206; behavioral 1,250 of 1,514 — one shard timed out; conformance 97 of 222 — timed out; conformance-dev/sandbox/diag/integration/htmx unmeasured). Bottom line: **the kernel itself conforms; the packaging and the test environments are where the defects are.**

## F-16 [S1, high, CONFIRMED] Shipped browser hyperscript cannot call functions: `host-call-fn` is referenced but defined nowhere in the shipped stack
- The shipped stack ships a full hs engine (`shared/static/wasm/sx/hs-{tokenizer,parser,compiler,runtime,integration,htmx,worker,prolog}.sx(+.sxbc)`), but `hs-runtime.sx(bc)`/`hs-integration.sx(bc)` reference `host-call-fn`/`host-call-fn-raising`, which only `run_tests.ml:3564` defines (the test runner's mock host bridge). The shipped platform (`sx-platform.js`) registers `host-call` / `host-new` / `host-get` etc. — NOT `host-call-fn`.
- Reproduced: running the behavioral corpus on the shipped stack as-is → **536 pass / 978 fail, 906 of them "Undefined symbol: host-call-fn"** (wasm-hs-behavioral.log). Any hyperscript that invokes a function or method would fail the same way in a real browser.
- The OCaml runner's own comment calls this binding "the single biggest gap: ~900 behavioral tests failed" — the fix was made in the test runner only, never in the shipped platform. Textbook case of F-7 (feature exists only in the runner's private environment).

## F-17 [S1, high, CONFIRMED] Shipped `hs-runtime.sx` is missing the `(jit-exclude! "hs-*")` guard
- `lib/hyperscript/runtime.sx` ends with `(jit-exclude! "hs-*")` + a comment explaining the hs recursion web **miscompiles under the bytecode JIT** (the parser-combinator JIT bug) and must stay CEK-interpreted.
- `shared/static/wasm/sx/hs-runtime.sx` is byte-identical EXCEPT this guard is absent (10-line diff; the other 5 hs modules are identical copies). The browser is precisely the sxbc/JIT-heavy environment. So the tested configuration (JIT-excluded) is not the shipped configuration (JIT-eligible). The lib→wasm/sx sync that copied these files dropped exactly the safety-critical line.

## F-18 [S2, high, CONFIRMED] The kernel conforms; the native runner's mock DOM is the outlier — and the "272 red" band mostly indicts the test env, not the code
Per-test diff, shipped-kernel+happy-dom (with the host bridge mirrored) vs native mock-DOM run:
- Pure pipeline (tokenizer/parser/compiler/runtime, 206 tests): **0 WASM-only failures**; all 51 failures shared with native; 2 native-only.
- Behavioral (1,250 matched): **9 WASM-only failures** vs **118 native-only failures** (WASM+happy-dom passes where the mock DOM fails: fetch 22, toggle 10, on 10, make 8, repeat 6, append 6, ...). Conformance (97 matched): 0 WASM-only, 4 native-only.
- Consequence for F-10: the permanently-red 272 hs failures on the canonical host are largely **mock-DOM environment artifacts** — the engine passes those tests on the shipped kernel with a more realistic DOM. The red band hides real information in both directions.
- The 9 WASM-only failures (listed in wasm-hs-compare.txt): behavior scoping ×4 ("Expected 10/20, got <empty>"), as-fragment conversions ×3 (a `{:__host_handle N}` leaks where a list is expected — happy-dom NodeList boundary), init/where sequencing ×2. Candidates for bisection; none look like core-semantics bugs.
- Also measured: **the js_of_ocaml kernel is ~1-2 orders of magnitude slower** on this corpus than native (conformance ≈24s/test vs seconds for the whole file natively) — worth knowing before making WASM corpus runs a gate.

## F-19 [S3, med, CONFIRMED] hs corpus drift + inverted assert labels
- The shared-failure bucket (~50 behavioral + parser/tokenizer suites) is corpus drift: generated tests (generate-sx-tests.py, "DO NOT EDIT") still expect the old parser AST — `(me)` implicit target and `(. obj k)` — while the current parser emits `(beingTold)` / `poss` (probed identically on native sx_server AND WASM: `(hs-compile "add .foo")` → `(add-class "foo" (beingTold))` on both). Tests were not regenerated after the parser change.
- `assert=` (spec/harness.sx:31) is `(actual expected msg)` but the generated corpus calls it as `(assert= expected actual)` — every failure message prints Expected/got **swapped**, which materially misled diagnosis during this probe (it makes current-parser output look like the "expected" value).
- `hypersx.sx` in the shipped boot list is NOT the hyperscript engine (it's an sx→hypersx-notation pretty-printer); the actual hs engine modules are shipped but absent from the boot module list (`loadWebStackFallback`, sx-platform.js:670) — load path for hs in production is unclear/on-demand only.

---

# SUSPECTED findings / gaps (probe proposed, not yet reproduced)

## S-1 [S1, med] Everything F-8 lists between native and the JS bundle likely also splits native vs WASM in web-stack .sx code paths not covered by my 130 probes (strings near render, regex at the sxbc layer, signals timing). Probe: run the full probe corpus + web adapter smoke through `run_wasm.js` and diff (harness exists; only 130 exprs run so far).

## S-2 — RESOLVED, promoted to F-16/F-17/F-18 below (hs corpus WAS run against the shipped WASM kernel; see "Hyperscript on the shipped kernel" section).

## S-3 [S2, med] `import`-dependent components behave differently in production browser vs test: OCaml runner resolves imports synchronously from disk; browser resolves via fetch + sxbc bundles (compile-modules.js). No test asserts the browser-side import path resolves the same module set. Probe: compare `sx_build_manifest`/dist sxbc contents vs run_tests preload list.

## S-4 [S3, med] Float printing (`0.3` vs `0.30000000000000004`) means any golden-string test containing floats validates different precision per host — some current "passes" may be precision-coincidences. Probe: grep corpus asserts for float literals with >6 significant digits.

## S-5 [S3, med — partially confirmed] The two JS kernel BUILDS disagree with each other: `spec/tests/test-adt.sx` (algebraic-data-types) passes on the standard bundle but fails 53 assertions on the `--full` build (`--extensions continuations --spec-modules types`). Confirmed from js-standard.log (0 FAIL, PASSes present) vs js-full.log (53 FAIL). So build flags change language behavior within one host — the types/continuations extension modules interfere with ADT/match. Suspected mechanism unverified; probe: bisect by building with each flag alone and running test-adt.sx.

---

# Suite asymmetry — why the counts differ (summary answer to the brief)

- Corpus exploded (hyperscript upstream ports, GQL, parser combinators, regex, records, chars, bytevectors, numeric tower, values/promises...) to ~5-6k assertions; briefed counts are months stale.
- JS standard (5086) = spec/tests only; JS full (5656) = + lib/tests + bigger kernel build.
- OCaml (6036) = spec/tests + web/tests + native foundation tests, but NOT lib/tests.
- Python = 0 (deleted).
- WASM (shipped browser artifact) = ~0 (boot smoke + 2 files).
- **No two hosts run the same file set; only spec/tests overlaps fully — and even there the runner environments differ (F-5, F-7), so "same test passing on two hosts" does not mean "same behavior".**

# Recommended conformance gate (one paragraph for the maintainer)
Run the spec/tests corpus on: native run_tests.exe, the WASM kernel via node (harness pattern: `scratchpad/probes/run_wasm.js`), and (if kept) the JS bundle — from the SAME preload manifest, with runner shims deleted or mirrored into all three; add lib/tests to the OCaml runner; skip-list the browser-only hs suites instead of letting them fail; add a numeric/string/dict-order differential probe file (seed: `scratchpad/probes/probes.txt`, 130 exprs, 32 divergences today) that must be output-identical across kernels; regen sx_ref.ml in CI and diff against checked-in.