vista-cloud-dev · rafael5 · Jun 14, 2026 · Jun 14, 2026 · Jun 14, 2026
diff --git a/dist/repo.meta.json b/dist/repo.meta.json
@@ -7,7 +7,7 @@
   "layer": "m",
   "license": "AGPL-3.0",
   "agent_instructions": "AGENTS.md",
-  "verified_on": "2026-06-13",
+  "verified_on": "2026-06-14",
   "exposes": {
     "modules": "dist/stdlib-manifest.json",
     "errors": "dist/errors.json"

diff --git a/dist/stdlib-manifest.json b/dist/stdlib-manifest.json
diff --git a/docs/memory/MEMORY.md b/docs/memory/MEMORY.md
@@ -2,6 +2,7 @@
 
 One line per memory file. Content lives in the files, not here.
 
+- [iris-native-backends](iris-native-backends.md) — PR #1: the 3 optional modules' IRIS dispatch arm uses the **inlined `$zversion["IRIS"` probe** (not a public engine helper — that part of the PR was dropped as superseded); dual-engine local-test runbook (**YDB needs `--chset m`**, rebuild `/tmp/m` for the flag); `m-test-iris` embedded-Python is non-functional so STDCOMPRESS-IRIS is unverifiable locally; how the stale PR was landed without merge/rebase/force-push (forward-commit-to-master-tree).
 - [waterline-g1-gate](waterline-g1-gate.md) — the m/v waterline **G1 gate** (`m arch check` in m-cli) — dependency-direction (v→m only); how `layer` is declared (dist/ meta vs root `repo.meta.json` for m-cli's gitignored dist/), check-manifest doesn't schema-validate the meta, and the v-cli registry-regen `go mod tidy` dep. Built s12 (loose end C).
 - [t0b2-msl-kids-base](t0b2-msl-kids-base.md) — VSL T0b.2 (MSL KIDS-install-as-green): **YDB leg GREEN — 15/15 test-in-place** after the m-ydb gbldir (`e5dcf85`) + v-pkg streamed-install (`aa1991f`) fixes. **IRIS leg (s9):** `raises^STDASSERT` **now ported to IRIS** (try/catch `irisRaises` branch; YDB byte-identical) → STDFMT/STDARGS clean + STDASSERTTST 40/40 both engines + STDUUID P2 gone; remaining IRIS crashes are **non-raises**. **(s10):** file I/O made dual-engine — STDFS portable facade (`$$openRead/Write/Append`+`readLn`) + STDOS.env IRIS arm + 5 consumers migrated; **STDFSTST 50/50 both engines, YDB full 2098/0**. **But the consumer SUITES still don't go green on IRIS** — separate non-file blockers (STDJSON **byte-mode** parser, STDCSV **`@cb@` indirection**, **wide-char** descriptions). file-I/O ≠ green suites; see §s10. Full 15/15 needs byte-mode + callback-idiom + wide-char work (out of file-I/O scope). ≤8-char-name decision keeps STDASSERT/STDSEMVER as a rename follow-up.
 - [vsl-doc-gaps-v0.2](vsl-doc-gaps-v0.2.md) — how the VistA Standard Library architecture doc's §12 VDL gaps resolved at v0.2; the vdocs `XU:XU:UG` over-collapse defect that blocks gold-promotion of the Kernel feature guides.

diff --git a/docs/memory/iris-native-backends.md b/docs/memory/iris-native-backends.md
@@ -0,0 +1,55 @@
+---
+name: iris-native-backends
+description: PR #1 reconciliation — the 3 optional modules' IRIS dispatch arm uses the inlined `$zversion["IRIS"` probe (not a public engine helper); dual-engine local-test runbook; m-test-iris embedded-Python gap; how a stale PR was landed without merge/rebase/force-push.
+metadata:
+  type: project
+---
+
+**IRIS-native backends for the 3 optional modules** (STDCRYPTO / STDCOMPRESS /
+STDHTTP) landed via **PR #1** ("B2"), reconciled onto the s9–s12 IRIS sweep on
+2026-06-14. The stale PR predated that sweep and was CONFLICTING + partially
+superseded; reconciliation kept only the novel payload.
+
+**The engine seam is the inlined `$zversion["IRIS"` probe — the house idiom
+for runtime modules.** Each module's dispatch helper gets an
+`if $zversion["IRIS" quit $$iris…(…)` arm ahead of the YDB `$&pkg.fn` path
+(STDCRYPTO → `$SYSTEM.Encryption.SHAHash/.HMACSHA`; STDHTTP →
+`%Net.HttpRequest`; STDCOMPRESS → embedded-Python zlib + ctypes/zstd). The
+PR's own `$$engine^STDOS()` helper was **dropped as superseded** — master's
+STDOS is already IRIS-ported by inlining the same probe, STDASSERT already has
+its `irisRaises` try/catch arm (s9), and `$$engine^STDHARN()` (internal) exists
+if a helper is ever wanted. **Future IRIS arms in m-stdlib runtime code: inline
+`$zversion["IRIS"`, do not add a cross-module engine helper.** See
+[[t0b2-msl-kids-base]] for the s9–s12 idioms ($ZTIMESTAMP clock, xecute-built
+dispatch, STDFS facade).
+
+**Dual-engine local-test runbook** (the canonical invocations — the Makefile
+targets omit these flags because they target host-YDB CI / the Python m-cli):
+- Build a current `m`: the committed `m-cli/dist/m` can be stale; rebuild with
+  `cd m-cli && GOPROXY=file://$HOME/go/pkg/mod/cache/download GOSUMDB=off GOFLAGS=-mod=mod go build -o /tmp/m .`
+  (the `--chset` flag was missing from the committed binary).
+- **YDB:** `m test tests/X.m --engine=ydb --docker=m-test-engine --routines src --chset m`
+  — **`--chset m` is mandatory**: the m-test-engine container defaults to
+  `ydb_chset=UTF-8`, but the byte/binary modules need byte mode (else
+  `%YDB-E-BADCHAR` on raw digest/compress bytes).
+- **IRIS:** `m test tests/X.m --engine=iris --docker=m-test-iris --routines src --namespace USER`
+  (byte mode is inherent on IRIS).
+- Results: STDCRYPTOTST 23/23, STDHTTPTST 68(YDB)/67(IRIS), STDCRYPTODOCTST 1/1,
+  STDCOMPRESSTST 59/59 (YDB) all green.
+
+**Gap — `m-test-iris` (iris-community image) has non-functional embedded
+Python**: `%SYS.Python` class exists but `Import("sys").version`→0, and the
+STDCOMPRESS IRIS path aborts non-trappably (0/0). So **STDCOMPRESS-IRIS can't
+be verified locally**; the PR validated it on `vista-iris` (working embedded
+Python). The reconciled code is the PR's vista-iris-validated logic with only
+the (proven-correct) seam changed. Discoveries register has the detail.
+
+**Landing a stale PR under this sandbox** (`git merge`, `git rebase`,
+`git clean`, `rm`, and force-push are all denied): don't rebase. Make a
+**forward commit on the branch** whose tree equals `master + additive
+backends` — `git checkout origin/master -- .` (sync whole tree to master),
+re-`git rm` master's deletions, `git checkout HEAD -- <the backend files>` to
+restore the wanted changes, regenerate `dist/`. GitHub's 3-way merge is then
+clean (master's changes appear identically on both sides; backends are purely
+additive in files master never touched) and the PR diff is minimal. Verify
+with `git diff --cached --stat origin/master` before committing.
diff --git a/docs/tracking/discoveries.md b/docs/tracking/discoveries.md
@@ -1,7 +1,7 @@
 ---
 created: 2026-05-10
-last_modified: 2026-06-07
-revisions: 3
+last_modified: 2026-06-14
+revisions: 4
 doc_type: [NOTES]
 ---
 
@@ -71,6 +71,9 @@ requires "no open P0/P1 entries against those subjects."
 | 2026-06-13 | P2 | m-stdlib | Consumer SUITES stay IRIS-red for **non-file** reasons (byte-mode · `@cb@` indirection · wide-char) | After the file-I/O facade, the consumer suites still 0/0 on IRIS — **the file I/O was only part of why.** Confirmed by isolated probes: **STDJSON** — `$$parse^STDJSON` (no file) crashes on IRIS = the **byte-mode** assumption (STDJSON/STDB64/STDHEX/STDCSPRNG treat 1 char == 1 byte; IRIS strings are 16-bit Unicode) — documented charset constraint; **STDCSV** — `do @callback@(rownum,.fields)` (indirection-with-args) crashes on IRIS even with ASCII (`parseFile` callback dispatch); **STDCSV/STDSEED/STDLOG/STDXML** descriptions also carry non-ASCII (wide-char output, the m-iris GetOut/session-capture lane). STDCSV's *parser core* (`$$parse` of a string) DOES pass on IRIS (2/2), so it's the callback + wide-char, not the parser. | Out of file-I/O scope — separate follow-ups: byte-mode portability (big; affects the byte-oriented modules), an IRIS-portable callback-dispatch idiom for `parseFile^STDCSV` (replace `@cb@(args)` with an `xecute`d call or a fixed dispatch), and the wide-char capture path. Tracked so the next session knows the file-I/O refactor alone does NOT green these suites. | **superseded 2026-06-13 (s11)** — the "STDJSON byte-mode" half was a MISDIAGNOSIS (see s11 row below); the `@cb@` + wide-char halves stand (wide-char now fixed). |
 | 2026-06-13 | P1 | m-stdlib | **RE-BASELINE (s11): T0b.2 IRIS leg is 10/15; the gap is 4 code fixes, NONE byte-mode** | Rebuilt `m-iris/dist/m-iris` from `m-iris-driver@49a5b00` (the GetOut wide-char fix; the prior dist binary predated it) and ran `kids-test-in-place.sh iris` on foia. **10/15** (was 6/15). **The GetOut wide-char fix works on remote** — STDURL 150/0, STDREGEX 102/0, STDFMT 62/0 all green now. **Crucially STDB64 (55/0) + STDHEX (49/0) — the byte family — PASS on IRIS**, so there is **NO byte-mode blocker**; the s10 "STDJSON byte-mode" claim was wrong (it never isolated the crash). The 5 reds, each a DISTINCT cause: **(1) STDJSON** crash `<SYNTAX> parse+12` for ALL inputs = the **unguarded `zgoto`-`$etrap`** idiom (parse + encode) — IRIS rejects YDB `zgoto LEVEL:label`; STDFS/STDHARN/STDASSERT.raises already guard the same idiom with an `if $zversion["IRIS" … quit` arm, STDJSON is the only base module that didn't. **(2) STDXML** crash `<SUBSCRIPT> parseElement+20 myNs("")` = **null/empty-string subscripts** (`myNs("")`/`nsMap("")` for the default namespace) — IRIS rejects null subscripts, YDB allows. **(3) STDCSV** won't COMPILE on IRIS: `<COMPILE> #5475 Expected end of line : '@callback@(curRow,.fields)'` — **IRIS has no ARGUMENT indirection** (`do @cb@(args)`), only name-indirection; whole routine fails → 0/0. **(4) STDDATE** 1 fail "year in plausible range": `now()` → `3567-05-6.157218T…` because it reads `$ZHOROLOG` as YDB's 4-comma `d,s,u,t` but **IRIS `$ZHOROLOG` is single elapsed-seconds**. **(5) STDUUID** 2 fails were COLLATERAL from the crashers in the same sequential process — source AND installed both 131/131 in isolation; no fix. | **4 code fixes (this session, TDD dual-engine, YDB byte-identical):** #1 STDJSON IRIS try/catch arm (parse+encode); #2 STDXML null-subscript fix [USER FORK → **sentinel-key in code**, chosen 2026-06-13] **[LANDED — `$$dfltNsKey()`=single space; STDXMLTST 209/209 both engines]**; #3 STDCSV `xecute`-built callback dispatch **[LANDED — + a latent STDFS readLn `$ECODE`-on-EOF bug; STDCSVTST 59/59 & STDFSTST 50/50 both engines]**; #4 STDDATE `$ZTIMESTAMP` IRIS arm in now() **[LANDED — STDDATETST 66/66 both engines]**. #5 STDUUID none. Then re-run on foia → expect 15/15. **#1 COMPLETE — STDJSONTST 209/209 BOTH engines (YDB + foia remote).** Beyond the etrap port: fixed TWO latent UTF-8 OPERATOR-PRECEDENCE bugs in emitUtf8 + the surrogate combine (M has no precedence — `$char(192+cp\\64)` evaluated as `(192+cp)\\64`; garbage on BOTH engines, latent because old tests used literal-byte passthrough); rewrote the 2 byte tests to `\u`-escapes+`$char`; graceful empty-key reject on IRIS (user decision — null local subscript is unconditional on IRIS incl. foia; documented in-code + stdjson.md + users-guide); fixed irisParse `$ECODE`-on-failure pollution (same class as STDFS readLn EOF); migrated STDJSONTST file tests off raw YDB `open:(newversion)` to the STDFS facade — **but STDJSONTST has 2 further IRIS tail issues exposed once the crash clears: (a) byte-exact UTF-8 tests use `$zchar` (unsupported on IRIS) + literal-multibyte source (byte-mode boundary) → needs the byte-mode decision; (b) `tParseObjectEmptyKeyAllowed` stores `root("")` = null subscript → same sentinel-key class as STDXML.** Lesson: each crashing suite stacks several IRIS issues; 10/15 undercounts per-suite work. | **CLOSED (s11): IRIS leg 17/15→17/17 on foia** (`suites=17 pass=1483 fail=0`). All 4 code fixes + STDUUID `unixMs` `$ztimestamp` IRIS arm (v7 time prefix was loosely-monotonic on IRIS via the YDB `$zhorolog` assumption) landed; STDFS+STDOS added to the base (STDOS ported to IRIS dual-engine first). YDB per-suite green; vehu in-place loop re-run owed. |
 | 2026-06-07 | P2 | vdocs | `consolidate` over-collapses the Kernel-8.0 per-feature User Guides into one anchor | Surfaced fetching the dedicated KIDS / Device Handler / TaskMan guides to fill the [VistA Standard Library architecture §12](../plans/msl-vsl-architecture.md) doc gaps. All ~41 distinct Kernel-8.0 `krn_8_0_{dg,sm}_*_ug` feature guides (KIDS, Device Handler, TaskMan, Alerts, Common Services, …) are assigned the **same `XU:XU:UG` anchor key** by the catalog/identity logic. `consolidate` treats them as one version group, keeps a single "winner", and demotes the other ~40 to `is_latest=0` — so they are fetched/converted/enriched/normalized and present in `index.db`, but **excluded from the FTS gold-search surface** (`vdocs ask` can't see them) and have no `documents/gold/consolidated/.../body.md` anchor. Same defect blocks `VIAB/via_vip_user_guide` (a separate fetch, also stuck at convert/promote). | m-stdlib side: the architecture doc's findings were read **directly from the normalized silver bodies** (`documents/silver/text/03-normalized/XU/<slug>/body.md`), so the doc does not depend on the fix. **vdocs side (open, upstream):** the `catalog`/identity stage must derive a per-document anchor key for granular feature guides (e.g. include `doc_subject`/slug, not just `doc_code=UG`), then re-run `consolidate`→`index`→`relate`→`manifest`. Until then these guides are cited by their fetched body paths and marked 🟡 gold-promotion-pending in the architecture doc §12/§13. | open (in vdocs) |
+| 2026-06-14 | docs | m-stdlib | PR #1 IRIS backends reconciled — engine seam is the inlined `$zversion["IRIS"` probe, not a public `$$engine^STDOS()` | Stale feature PR #1 ("B2: IRIS-native backends") predated the s9–s12 IRIS sweep. It added its own `$$engine^STDOS()` engine-detect helper + a STDASSERT `irisCapture` try/catch arm. **Both were superseded:** master's STDOS is already IRIS-ported (inlines `$zversion["IRIS"` per function, no public engine helper) and STDASSERT already has the `irisRaises` try/catch arm (s9). Reconciliation **dropped** the PR's STDOS + STDASSERT changes and **rewired** the 3 backends' 6 dispatch call-sites from `$$engine^STDOS()="iris"` → `if $zversion["IRIS"` (master's established runtime idiom, used in STDOS + STDASSERT + STDHARN.engine). The `$ZCHAR`→`$CHAR` change in STDCOMPRESSTST was kept (IRIS has no `$ZCHAR`; `$CHAR`≡`$ZCHAR` for 0..255 under `ydb_chset=M`, matching the rest of the byte suites). | Branch made mergeable without merge/rebase/force-push (all sandbox-denied): a forward commit set the branch tree to `master + additive backends` (revert STDOS/STDASSERT to master, regenerate dist), so GitHub's 3-way merge is clean. PR diff is now just the 3 backend modules + STDCOMPRESSTST + their manifest entries. | resolved (PR #1) |
+| 2026-06-14 | P3 | m-stdlib | `m-test-iris` (iris-community image) has **non-functional embedded Python** → STDCOMPRESS-IRIS unverifiable locally | The `intersystemsdc/iris-community:latest` container used for `m test --docker=m-test-iris` ships the `%SYS.Python` class (exists=1) but its embedded-Python runtime is not wired up: `##class(%SYS.Python).Import("sys").version` returns `0`, and calling the STDCOMPRESS IRIS path (`irisInit`→`b.exec`→helper) aborts the suite non-trappably (0/0). STDCRYPTO-IRIS (23/23) and STDHTTP-IRIS (67/67) verify fine — they use built-in classes (`$SYSTEM.Encryption` / `%Net.HttpRequest`), no Python. STDCOMPRESS-YDB is green (59/59 via libz/libzstd callout). The PR's original STDCOMPRESS-IRIS 59/59 was on `vista-iris` (working embedded Python). | The reconciled STDCOMPRESS IRIS logic is the PR's vista-iris-validated code, unchanged except the proven-correct seam — so the gap is the local engine image, not the code. Verify STDCOMPRESS-IRIS on a `vista-iris`-class instance (or once embedded Python is enabled in m-test-iris). Local-test runbook: `--docker=m-test-engine --routines src --chset m` (YDB byte mode — the container defaults to UTF-8) / `--docker=m-test-iris --routines src --namespace USER` (IRIS). Robustness follow-up (out of reconciliation scope): make the STDCOMPRESS IRIS path fail gracefully (return "FAIL"/0) when embedded Python is absent instead of a non-trappable abort. | open (env / follow-up) |
+| 2026-06-14 | P2 | tree-sitter-m | pinned grammar (CI's `m-dev-tools/tree-sitter-m`) can't parse `^||` process-private globals | STDCOMPRESS's IRIS backend memoized its embedded-Python init with a `^||STDCOMPRESS("py")` process-private global. The local Go `m` (newer bundled grammar) parses it fine, but CI's Python m-cli + pinned tree-sitter-m emits **2 ERROR nodes** on the two `^||` references → `m fmt --check` fails (`source did not parse cleanly`), red-gating the YDB job. (master uses no `^||`, so master is green.) | m-stdlib side: **dropped the `^||` memo** — `irisInit^STDCOMPRESS` now re-execs the Python defs per call (idempotent; `ctypes.CDLL` refcounts the same .so, redefining helpers is free). tree-sitter-m side (upstream, open): add `^||` process-private-global syntax to the grammar so the memo idiom is available to M code. | resolved (m-stdlib); open (tree-sitter-m lane) |
 
 ## Cross-references
 

diff --git a/docs/tracking/module-tracker.md b/docs/tracking/module-tracker.md
@@ -114,16 +114,25 @@ current state.
 | [x] | P4 | L25 | 27 | [`STDXML`](../modules/stdxml.md) | `v0.4.0` | 14d | none (completed) | none | XML 1.0 + Namespaces 1.0 + XPath 1.0 + DTD envelope | n/a |
 | [x] | P4 | L26 | 28 | [`STDMATH`](../modules/stdmath.md) | `v0.4.0` | 1d | none (completed) | none | Numeric helpers — clamp / min / max / sum / count / mean | n/a |
 | [x] | P4 | L27 | 29 | [`STDXFRM`](../modules/stdxfrm.md) | `v0.4.0` | 1d | none (completed) | none | Higher-order array transforms — map / filter / reduce | n/a |
-| [x] | P3 | H1 | 30 | [`STDCRYPTO`](../modules/stdcrypto.md) | `v0.4.0` | 2d | none (completed) | `$&stdcrypto.fn → libcrypto`; A6 | SHA-256/384/512 + HMAC-SHA-256/384/512 | 🟡 C12 |
-| [x] | P3 | H2 | 31 | [`STDCOMPRESS`](../modules/stdcompress.md) | `v0.4.0` | 6d | none (completed) | `$&stdcompress.fn → libz + libzstd`; A6 | gzip / gunzip / deflate / inflate / zstdCompress / zstdDecompress | 🟡 C13 |
-| [x] | P3 | H3 | 32 | [`STDHTTP`](../modules/stdhttp.md) | `v0.4.0` | 4d | none (options) | STDURL; `$&stdhttp.fn → libcurl`; A6 | HTTP/1.1 client + pure-M wire-format helpers | 🟡 C14 |
+| [x] | P3 | H1 | 30 | [`STDCRYPTO`](../modules/stdcrypto.md) | `v0.4.0` | 2d | none (completed) | `$&stdcrypto.fn → libcrypto`; A6 | SHA-256/384/512 + HMAC-SHA-256/384/512 (+ IRIS-native arm: `$SYSTEM.Encryption`) | 🟡 C12 |
+| [x] | P3 | H2 | 31 | [`STDCOMPRESS`](../modules/stdcompress.md) | `v0.4.0` | 6d | none (completed) | `$&stdcompress.fn → libz + libzstd`; A6 | gzip / gunzip / deflate / inflate / zstdCompress / zstdDecompress (+ IRIS-native arm: embedded Python zlib+ctypes/zstd) | 🟡 C13 |
+| [x] | P3 | H3 | 32 | [`STDHTTP`](../modules/stdhttp.md) | `v0.4.0` | 4d | none (options) | STDURL; `$&stdhttp.fn → libcurl`; A6 | HTTP/1.1 client + pure-M wire-format helpers (+ IRIS-native arm: `%Net.HttpRequest`) | 🟡 C14 |
 | [ ] | — | T1 | 33 | [`STDHARN`](../modules/stdharn.md) | — | 3d | P2 `^%MONLBL` coverage (STDCOV) · P4 STDWATCH hooks | STDASSERT (no-halt orchestration mode) | Resident pure-M test/coverage harness — frames `^STDASSERT` suites for m-cli 5.1 (server-side delegation) | ✅ `internal/harness` (P0–P1) |
 
 **Aggregate.** ~108d shipped across all 32 landed modules (sum of
 the Effort column above). **Full engine suite green on `main`
 2026-05-08: 32 suites, 2483/2483 assertions.** All three Phase 3
 modules engine-green: STDCRYPTO H1 (23/23), STDCOMPRESS H2 (59/59),
-STDHTTP H3 (68/68). **All numbered tickets T1–T30 closed.** Optional
+STDHTTP H3 (68/68). **IRIS-native backends landed (PR #1, 2026-06-14,
+reconciled onto the s9–s12 IRIS sweep):** each optional module gained
+an `if $zversion["IRIS"` arm in its dispatch helper (STDCRYPTO →
+`$SYSTEM.Encryption`, STDHTTP → `%Net.HttpRequest`, STDCOMPRESS →
+embedded-Python zlib+zstd) — the engine seam is the inlined `$zversion`
+probe (master's idiom), **not** a public `$$engine^STDOS()` helper (that
+part of the PR was dropped as superseded). Dual-engine verified locally:
+STDCRYPTO 23/23 and STDHTTP 67/67 green on IRIS (m-test-iris) as well as
+YDB; STDCOMPRESS-IRIS needs working embedded Python (the iris-community
+image lacks it — see discoveries 2026-06-14). **All numbered tickets T1–T30 closed.** Optional
 add-ons (rows tagged `none (options)`: T15 / T16 / T17 / T18 / T19 /
 T22 / STDHTTP iter 3) sit behind concrete-consumer drivers and are
 not gating any release. Per-module deep history (scaffolding, migrations,