release: cut v0.33.0 — the silent-wrong-answer release

CLAUDE.md

@@ -146,28 +146,30 @@ bash tools/embed_stack_soak.sh  # embed REPL soak inside a 64 KiB stack rlimit (
 
 ## Current state & where the detail lives
 
-- **Latest release: v0.32.0** (2026-07-16) — the desktop-shell release:
-  lib/ui's gap series (#561–#577, #594), surfaced by DeslanStudio's port —
-  the toolkit's first real consumer — and closed. The toolkit now owns what
-  the shell hand-rolled: `handle_key` (#563, keyboard through `dispatch`
-  instead of trapped inside `app_loop`, priority pinned in one place),
-  `request_quit` (#564, ending the loop from a callback instead of `exit of
-  N` skipping `gfx_close`), clickable combobox dropdown items (#577 —
-  hit-test before closing, select by the click's own y, and exempt popups
-  from double-click detection), self-measuring `label` w/h + `auto_size`
-  (#561), `grid.owns_cells` + a configurable row-label gutter (#572),
-  piano-key release wherever the pointer lands (#570), hover feedback on
-  checkbox/toggle (#594), and **`menu_bar`** (#565 — owning its popup's
-  z-order via the `_render_popups` overlay pass + `_find_open_popup_at`,
-  since rendering above the tree is only half of z-order). Three of the ten
-  filed issues had already been fixed and were closed with probe output, not
-  re-fixed. (v0.31.0, 2026-07-14: DAW audio-I/O + live-synthesis — bulk
-  audio-I/O kernels #602/#603, the live audio-streaming primitive
-  `audio_stream_*` #612, `waveform_view` selection markers #610/#611.
-  v0.30.0, 2026-07-13: debugging-and-distribution + first-consumer —
-  `--step` tape stepper #418, tape format v2 #539, error carets/LSP ranges
-  #407, `--bundle` #413, entropy-walk cycle detection #571, gfx audio
-  capture #579 + buffer playback #578, the lib/ui input trio #567–#569.)
+- **Latest release: v0.33.0** (2026-07-24) — the silent-wrong-answer release,
+  plus the primitives the app fleet forced. A sweep of the plausible-but-wrong-
+  at-rc=0 class: `regex_match` dropping every capture after a non-participating
+  group (#629), `json_decode` truncating long numbers by 50 orders of magnitude
+  (#628), four stdlib math functions returning confidently wrong values
+  (#638–#641), a module's top-level assignment rebinding its *importer's*
+  global (#673), a comprehension var and a same-named local splitting into two
+  bindings so later program text changed an earlier statement's result
+  (#633/#642), `report_value` calling a converging sequence `diverging` (#674),
+  and an observer LR gate that measured 100% "stable" while the loss was still
+  falling. Primitives, each requested by a consumer that hit the wall: four
+  DEFLATE codecs behind `make zlib` (#684 — `.xlsx`/`.gz` readable from
+  script), tape-captured `clock_unix` (#683), and C-backed `list_index_of` /
+  `list_contains` / `list_insert_at` / `list_slice` (#543/#544). Plus
+  `kill -USR1` live observer dumps (#660) and a test runner that fingerprints
+  its own binary mid-run, bounds every block against a runaway, and derives its
+  tally from block output after the hand-synced literals were caught
+  under-reporting by 169 asserts (#654). (v0.32.0, 2026-07-16: the
+  desktop-shell release — lib/ui's gap series #561–#577/#594, surfaced by
+  DeslanStudio's port and closed, incl. `menu_bar` #565 and `handle_key` #563.
+  v0.31.0, 2026-07-14: DAW audio-I/O + live-synthesis — bulk audio-I/O kernels
+  #602/#603, `audio_stream_*` #612, `waveform_view` markers #610/#611.
+  v0.30.0, 2026-07-13: debugging-and-distribution — `--step` tape stepper #418,
+  tape format v2 #539, error carets/LSP ranges #407, `--bundle` #413.)
   Unreleased
   work on `main`: see CHANGELOG.md `[Unreleased]`. Full version history:
   **CHANGELOG.md** (don't re-narrate it here — tools/doc_drift_check.sh

ROADMAP.md

@@ -1,8 +1,8 @@
 # Roadmap
 
-Current version: **0.32.0**
+Current version: **0.33.0**
 
-Recently shipped (0.28.0 → 0.32.0, 2026-07-08 → 07-16):
+Recently shipped (0.28.0 → 0.33.0, 2026-07-08 → 07-24):
 
 - **0.29.0** — the strategic headline landed: the **deterministic
   cooperative task layer on the tape** (#408, seeded scheduler, zero
@@ -22,6 +22,13 @@ Recently shipped (0.28.0 → 0.32.0, 2026-07-08 → 07-16):
   #594) surfaced by DeslanStudio's port and closed, including `menu_bar`
   (#565), `handle_key`/`request_quit` (#563/#564), self-measuring `label`
   (#561), and `dialog` children + `file_dialog` (#575).
+- **0.33.0** — the silent-wrong-answer release plus the primitives the app
+  fleet forced: the plausible-but-wrong-at-rc=0 sweep (`regex_match` #629,
+  `json_decode` #628, four stdlib math bugs #638–#641, import top-level scope
+  #673, the listcomp/local double binding #633/#642, `report_value` #674), the
+  DEFLATE codecs behind `make zlib` (#684), tape-captured `clock_unix` (#683),
+  four C-backed list operations (#543/#544), `kill -USR1` live observer dumps
+  (#660), and the test-runner integrity pass (#681/#648/#654).
 
 Full per-version detail lives in [CHANGELOG.md](CHANGELOG.md) — this file
 is forward-looking.

VERSION

@@ -1 +1 @@
-0.32.0
+0.33.0

docs/llms.txt

@@ -1,6 +1,6 @@
 # EigenScript — single-file reference for language models
 
-> EigenScript v0.32.0. A small dynamically-typed language with one distinctive
+> EigenScript v0.33.0. A small dynamically-typed language with one distinctive
 > feature: an **observer** tracks every value's entropy/trend on each
 > assignment, so convergence, stability, and oscillation are first-class. This
 > file is the whole surface you need to generate correct `.eigs`. Models

tests/test_vm_run_bytecode.eigs

@@ -1,7 +1,14 @@
 # vm_run_bytecode — the self-hosting bootstrap bridge. Assemble a chunk from
 # EigenScript values (code bytes + constant pool) and run it on the C VM.
-# Opcodes/operands are exactly what compile_ast emits (16-bit operands are
-# little-endian); this is what an EigenScript-written compiler will produce.
+# Opcodes/operands are exactly what compile_ast emits (operands are
+# little-endian, 16-bit except OP_LINE's, which is 32-bit since #630); this is
+# what an EigenScript-written compiler will produce.
+#
+# This file is the ABI contract for external bytecode producers (ouroboros'
+# self-hosted codegen, iLambdaAi's vendored copy of it). Changing an opcode's
+# NUMBER or its OPERAND WIDTH breaks every one of them silently — a misaligned
+# chunk runs as garbage at rc=0 — and their CI cannot see it, because they pin
+# a release. So any such change belongs here, in the same PR (#704).
 
 load_file of "lib/test.eigs"
 
@@ -153,4 +160,25 @@ catch es348:
 assert_eq of [sl348, 1, "SET_LOCAL out-of-range slot raises (#348)"]
 assert_eq of [vm_run_bytecode of [[GET_LOCAL,255,255, RETURN], []], null, "GET_LOCAL out-of-range reads null (missing-param semantics)"]
 
+# --- #630/#704: OP_LINE's operand is 32-bit, and that width IS the ABI ---
+# #630 widened this operand from 16 to 32 bits. No opcode number changed, so
+# nothing here failed — this file emitted no OP_LINE at all — and the break
+# only surfaced at the next consumer bump, where ouroboros produced empty
+# output for every program and iLambdaAi's grader lost its top rung. Both were
+# silent: rc=0, no error, just a chunk executing misaligned garbage.
+LINE is 68
+# well-formed: OP_LINE <line:32> then 2 + 3. A 4-byte operand keeps the stream
+# aligned, so this is 5. Emit three bytes instead and ADD lands on the wrong
+# byte — which is exactly what every external producer did on the v0.33.0 bump.
+assert_eq of [vm_run_bytecode of [[LINE,7,0,0,0, CONST,0,0, CONST,1,0, ADD, RETURN], [2, 3]], 5, "OP_LINE 32-bit operand keeps the stream aligned"]
+# the pre-#630 shape — OP_LINE + 2 operand bytes — must NOT still evaluate to 5.
+# It is either verifier-rejected or misaligned garbage; either way the contract
+# is that it is no longer the arithmetic the producer intended.
+old_shape is vm_run_bytecode of [[LINE,7,0, CONST,0,0, CONST,1,0, ADD, RETURN], [2, 3]]
+assert of [old_shape != 5, "pre-#630 16-bit OP_LINE emission no longer silently works"]
+# a line number past 65535 — the regime #630 exists for — survives the operand.
+assert_eq of [vm_run_bytecode of [[LINE,161,134,1,0, CONST,0,0, RETURN], [99]], 99, "OP_LINE carries a line past 65535 (#630)"]
+# truncated 32-bit operand is rejected by the verifier, not read off the end.
+assert_eq of [vm_run_bytecode of [[LINE,7,0,0], []], null, "verify: truncated OP_LINE operand rejected"]
+
 test_summary of null