Port clojure.core/destructure into jolt-core/clojure/core/00-syntax.clj
and drop core-destructure (plus the d-* helpers) from core.janet. The
expander is now self-hosted, shrinking the seed per the jolt-1j0 epic.
It's def+fn* rather than defn because defn isn't defined yet at that
point in the syntax tier. Only the let macro consumes its output, so let
now splices [~@(destructure bindings)] to keep a tuple binding form.
Also fix a gap the old Janet version papered over via Janet's
keyword->string: :keys accepts keyword elements ({:keys [:major :minor]}),
so use name/namespace for the local + ns instead of str (which keeps the
colon). This was breaking sci's impl/namespaces.cljc. Added spec cases.
fn now desugars destructuring params like Clojure's maybe-destructured: each
non-symbol param becomes a gensym and the body is wrapped in a let that rebinds the
pattern, so fn* only ever sees plain params and the COMPILER handles it (it rejected
patterns before, falling back to the interpreter). loop follows Clojure too: gensym
one loop var per binding, loop* over those, destructure via an inner let, with an
outer let so later inits see earlier destructured names — recur arity stays correct.
Two representation gotchas: build the param/binding vectors via [~@..] so they're
tuple forms (conj yields a pvec the analyzer rejects), and use (symbol (str
(gensym))) since a bare (gensym) in an overlay macro body is a Janet symbol the
destructurer rejects.
Closes the fn/loop destructuring gaps. conformance 228x3, fixpoint, clojure-test-
suite 3930, full suite green, bench flat.
lazy-seq wraps its body in (make-lazy-seq (fn* [] (coll->cells (do ...)))); lazy-cat
wraps each coll in lazy-seq and concats (concat is already lazy). Both user-facing,
in 30-macros. Also drop two now-stale comments. Laziness preserved (lazy-seqs-spec
20/20, including infinite/self-referential).
conformance 228x3, fixpoint, clojure-test-suite 3930, full suite green.
defprotocol/defrecord/extend-type/extend-protocol/reify + the extend/proxy/
definterface stubs move to 30-macros (user-facing, not used by the compiler). They
emit Jolt's protocol/type special forms (protocol-dispatch/register-method/
make-reified/deftype).
The one subtlety: a protocol value is a :jolt/type-tagged struct, and the
interpreter can't tell an embedded opaque value from a tagged map literal being
constructed. So defprotocol builds it via a make-protocol fn call (exposed from
core) instead of an embedded literal — a fn result evaluates normally and even
compiles. reify emits a {kw (fn* ...)} map literal that make-reified evaluates
(build-eval-map yields a struct it can iterate, unlike a hash-map phm). defrecord
vecs its spliced field-let bindings (a lazy mapcat seq won't splice) and uses an
explicit fresh-sym for the map-> param (auto-gensym doesn't cross nested
syntax-quotes).
protocols-spec 21/21, conformance 228x3, fixpoint, clojure-test-suite 3930, full
suite green.
defn drops an optional leading docstring/attr-map then emits (def name (fn* ...)).
Single- and multi-arity both reduce to (fn* ~@body) so no arity branching is needed.
map? is true for symbol forms in Jolt, so the attr-map strip is guarded with symbol?.
defn- delegates to defn (privacy isn't enforced, as in Clojure's own defn-). Placed
before fresh-sym, which is itself a defn- now.
All five fundamental macros (fn/let/loop/defn/defn-) are now in the overlay.
conformance 228x3, fixpoint, clojure-test-suite 3930, full suite green.
let -> let* with destructuring pre-expanded via destructure (now exposed as a
clojure.core fn, which it is in Clojure too) so the compiler sees plain bindings —
analyze-bindings rejects patterns as uncompilable. loop -> loop* with raw bindings,
matching the prior Janet macro: loop can't pre-destructure without breaking recur
arity, so the interpreter handles pattern loops and the compiler falls back.
conformance 228x3, fixpoint, clojure-test-suite 3930.
fn -> fn* is a one-line head-swap (the analyzer treats fn* as the primitive). It's
in 00-syntax since the analyzer, kernel, and other overlay macro bodies all use fn.
First of the fundamental macros to move.
conformance 228x3, fixpoint, clojure-test-suite 3930.
Port of core-for: desugar a comprehension to nested map/mapcat over the binding
colls, :let -> let*, :while -> take-while on the coll. Lives in 00-syntax because
doseq (already there) expands to it; the macro body uses only kernel/seed fns so it
runs at analyzer-build time. doseq no longer depends on a Janet macro.
Fixed a latent bug the Janet macro had: :when wrapped the inner form in (list ...)
unconditionally, but for an outer binding group the inner form is already a seq, so
mapcat produced a seq-of-seqs instead of flattening. e.g.
(for [x [0 1] :when (odd? x) y [:a :b]] [x y])
gave ((... ...)) instead of ([1 :a] [1 :b]). Now the (list ...) wrap is only applied
to the last group's scalar body; outer groups contribute their seq directly. Added
:let+:when, multi-group :when, and destructuring regression cases.
conformance 228x3, fixpoint, clojure-test-suite 3930, full suite green, bench flat.
Same shortcut as the prior Janet macro — realize a for comprehension with count
for side effects, return nil — so for keeps handling :when/:let/:while and multiple
bindings. Lives in 00-syntax because the analyzer uses doseq (analyze-try). Added
:when/:while/:let/returns-nil regression cases.
conformance 228x3, fixpoint, clojure-test-suite 3930, full suite green.
The real bug behind the case/cond-> fixpoint regression wasn't gensym — it was
build ordering. A top-level defn in a pre-kernel tier (the fresh-sym helper these
macros need) gets compiled in compile mode, and that lazily builds the self-hosted
analyzer via ensure-analyzer. 00-syntax loads before 00-kernel, so the analyzer
was built against a core missing mapv/second/peek/...: its references to them were
interned as forward-ref nil cells in jolt.analyzer. Those nil cells then shadowed
the real clojure.core defs when the analyzer rebuilt itself (stage2), so the
analyzer's own mapv calls went to nil — 'Cannot call nil as a function'. Only
variadic-heavy paths (juxt+mapv) surfaced it.
build-compiler! already documented that the kernel must be loaded first; nothing
enforced it. Gate ensure-analyzer on a :kernel-ready? flag set after the kernel
tier loads. A pre-kernel compile now falls back to the interpreter (compile-and-eval
already handles that) instead of building the analyzer too early. fresh-sym ends up
interpreted during 00-syntax load, which is fine.
With ordering fixed, case and cond-> move to 00-syntax (they need a real gensym, so
syntax-quote auto-gensym alone won't do). conformance 228x3, fixpoint stage1==2==3,
clojure-test-suite 3930, full suite green, bench flat.
Threading macros (recursive; the expand-once cache makes that free) and declare
(a no-op on Jolt — forward refs resolve via pending cells). They live in 00-syntax
because the analyzer itself uses -> and declare; validated by conformance 228x3
(the bootstrap-compiled analyzer expands them).
conformance 228/228 x3, clojure-test-suite 3930.
The hot control macros now live in the Clojure syntax tier. This was blocked
before: as interpreted overlay macros they re-expanded on every eval, timing out
a battery file (3930->3911). With the expand-once macro cache (prior commit) they
expand a single time with zero runtime cost, so moving them is free.
conformance 228/228 x3, clojure-test-suite 3930 (9 timeouts, not 10), full suite
green, bench flat (overlay vs janet macros within noise).
The interpreter re-expanded a macro call on every evaluation — so a macro in a
hot loop re-expanded each iteration (the runtime cost, and why moving the hot
macros to the overlay timed out battery files). Now each macro CALL form expands
once, keyed by form identity, and the macro-free expansion is reused; this is the
proper Lisp model (macroexpansion is a compile-time step). Also gives compile-once
gensym semantics — a foo# auto-gensym is now fixed across calls instead of fresh
per re-expansion. Cache cleared when a macro is (re)defined.
conformance 228/228 x3, full suite green.
A defmacro now compiles its expander to a native Janet fn — built as
(fn* args body...) and run through the self-hosted pipeline (its backtick body is
now compilable, prior commit) — instead of an interpreted closure. Macro
expansion happens at compile time with zero runtime cost, the proper Lisp model.
Wiring: evaluator exposes a macro-compile-hook the api sets to backend/
try-compile-fn; the defmacro handler uses the compiled expander when available,
falling back to the interpreted closure when the analyzer isn't built yet (early
tiers) or the body uses &env/&form (no such params on the compiled fn).
conformance 228/228 x3 (incl. REPL defmacro + gensym hygiene), full suite green.
Following the canonical read->macroexpand->compile model (Clojure LispReader /
tools.reader): a syntax-quote is lowered to plain construction code instead of
being interpreted. Adds form builders (__sqcat -> array, __sqvec -> tuple,
__sqmap, __sq1) and syntax-quote-lower (mirrors the interpreter's syntax-quote*/
sq-symbol exactly: foo# auto-gensym, core/special unqualified, else qualify;
~ -> expr, ~@ -> splice). The analyzer now handles syntax-quote as a special:
lower then analyze the result. The interpreter keeps the self-contained
syntax-quote* (no core dependency, works in a bare ctx); the two are cross-checked
by conformance interpret-vs-compile.
Effect: a backtick body compiles (1.96s -> 1.19s/200k) instead of falling back to
the interpreter. Next: compile defmacro expanders for the full macro speedup.
conformance 228/228 x3, full suite green.
New 00-syntax tier loaded FIRST (before 00-kernel), interpreted defmacros, so the
control macros the compiler and every later tier depend on can live in Clojure.
Validated by moving when: the kernel tier, self-hosted analyzer and seq/coll
tiers all compile against the overlay when. Constraint: syntax-tier macros may use
only special forms + core-renames seed primitives (not second/peek/kernel fns).
conformance 228/228 x3, full suite green.
Phase 3 batch 6 (jolt-1j0). binding installs an array-map var->value thread
frame and restores it on exit via try/finally. Dynamic rebinding is seen by
called fns. This completes the cleanly-portable safe macros (19 total).
conformance 228/228 x3, full suite green.
Phase 3 batch 5 (jolt-1j0). condp with a recursive emit building the nested if
chain, including the test :>> result-fn form and the trailing default. This
exhausts the cleanly-portable safe macros.
conformance 228/228 x3, full suite green.
Phase 3 batch 4 (jolt-1j0), 5 macros. cond->> (thread-last) is safe; cond->
stays (compiler uses it). delay/future expand to make-delay/future-call host
fns; letfn builds its fn* binding via a template (cons/list in a macro body make
a plist the evaluator can't call as a form).
conformance 228/228 x3, clojure-test-suite 3930, full suite green, bench flat.
Phase 3 batch 3 (jolt-1j0), 5 macros. Establishes the macro-body toolkit:
fresh-sym = (symbol (str (gensym))) for a shared explicit jolt symbol (a bare
(gensym) returns a Janet symbol the destructurer rejects), and splicing binding
pairs into a TEMPLATE vector so core-let sees a tuple form, not a runtime pvec.
when-first stays first-based (lazy-safe): Clojure's (seq coll) form realizes an
infinite coll like (repeat nil) under Jolt's eager evaluator and hangs — the
per-batch battery caught it (when_first.cljc). Clojure-correct seq form waits on
Phase 5 laziness.
conformance 228/228 x3, clojure-test-suite 3930, full suite green, bench flat.
Phase 3 batch 2 (jolt-1j0), 5 user-facing macros as defmacro + syntax-quote.
The conditional-binding macros use the auto-gensym temp# idiom so the name binds
only in the taken branch (the else/empty branch sees the outer scope, carrying
forward the earlier scope fix). when-let stays in Janet for now — 20-coll uses
it, so it must remain available before the macro tier loads.
Note: gensym in a macro body resolves to Janet's 0-arity builtin, so explicit
(gensym "prefix") fails — use template auto-gensym (foo#) instead.
conformance 228/228 x3, full suite green (incl. the if-let scope regression spec,
now exercising the overlay macro), clojure-test-suite 3930, bench A/B flat.
New 30-macros tier (registered after 20-coll) for user-facing macros expressed as
defmacro + syntax-quote instead of hand-built Janet form-transformers. Validated
end to end: overlay-defined macros expand in interpret, compile AND self-host
modes (conformance 228/228 x3).
Moved comment and if-not; removed their Janet core-X fns + core-macro-names
entries. Note: Jolt defmacro is single-arity, so multi-arglist macros become one
arglist with & rest / destructuring (if-not uses [test then & [else]]). Macros
used by the compiler or earlier tiers (and/or/when/cond/case/cond->/->/declare/
doseq/when-let) stay in Janet until a load-order story exists for them.
full suite green, clojure-test-suite 3930.
Eighth pure-fn batch (jolt-1j0 phase 2), 2 leaf fns. dedupe is eager consecutive
dedup (no transducer arity, as before); seq-to-map-for-destructuring is the
internal &{:keys} helper. Largely exhausts the easy pure-eager tier.
conformance 228/228 x3, full suite green (incl. destructuring).
Sixth pure-fn batch (jolt-1j0 phase 2), 3 leaf fns, now with regression tests
per the per-batch workflow. reductions is the canonical form, so (reductions f [])
calls (f) -> [0] instead of the prior []; tree-seq is eager pre-order DFS.
conformance 228/228 x3, clojure-test-suite 3929, full suite green, bench A/B flat.
Spec suites (interpret) for the overlay-migrated fns and both fixes:
- control-flow: if-let/when-let/if-some/when-some else-scope (9 cases)
- predicates: not-any?/idents/numeric preds/NaN?/abs/object?/... (24 cases)
- sequences: nthrest()/nthnext/distinct? value-eq/replace nil/take-last/... (23)
Plus 10 cases added to the all-3-modes conformance set for the highest-risk
fixes (if-let scope, nthrest (), distinct? value-eq), so compile + self-host are
guarded too: conformance 218 -> 228 in all three modes.
Full suite green.
Two correctness bugs found while migrating:
- nthrest returned nil where Clojure returns () : for n>0 the walk yields
(seq xs) wrapped in (or ... ()), so an exhausted/nil walk is () not nil
(only n<=0 returns coll). Both the prior Janet impl and the first overlay
port got this wrong. nthrest.cljc 13/1/0 -> 14/0/0.
- if-let/when-let/if-some/when-some leaked their binding into the else branch:
they wrapped the whole if in (let* [name val] ...), so (let [x 5] (if-let
[x nil] x x)) returned nil instead of 5. Fixed to bind a fresh temp around the
if and rebind the name only inside the taken branch, matching Clojure.
conformance 218/218 x3, clojure-test-suite 3928 -> 3929.
Fifth pure-fn batch (jolt-1j0 phase 2), 6 leaf fns. nthrest uses the canonical
loop form (same neutral 13/1/0 on nthrest.cljc as the prior Janet version; the
1 fail is a pre-existing platform-conditional case). object?/undefined? are
always false on Jolt.
conformance 218/218 x3, clojure-test-suite 3928, core-bench A/B noise-only.
Fourth pure-fn batch (jolt-1j0 phase 2), 6 leaf fns. distinct? now uses
value-semantics (the prior Janet impl keyed a table by identity, so equal
collections compared distinct); nthnext uses the canonical loop form, which
fixes (nthnext nil nil) => nil and the nil-count cases the prior pre-check threw
on. replace preserves nil values via get-with-default.
conformance 218/218 x3, clojure-test-suite 3927 -> 3928 (canonical nthnext),
core-bench A/B flat. Per-batch gate caught a transient -1 from the first
nthnext rewrite; fixed before commit.
Second pure-fn batch (jolt-1j0 phase 2). All leaf fns; canonical Clojure defs.
split-with now returns seqs for the parts (drop-while is a seq), matching
Clojure rather than the prior all-vector result — value-equal either way.
conformance 218/218 x3, clojure-test-suite 3927, core-bench 2333ms (~baseline).
First pure-fn migration batch. Both are leaf fns (no internal Janet callers);
the Clojure defs (not . some / not . every?) match the prior Janet behavior.
Removed the core-* defns and their core-bindings entries.
conformance 218/218 x3, clojure-test-suite 3927, core-bench 2340ms vs 2336ms
baseline (noise).
Classify the 421 core-* fns into seed/macro/host/lazy/movable buckets and record
a compile-mode benchmark baseline to catch regressions as fns move from native
Janet to the self-hosted overlay. Worklist in jolt-core/clojure/core/MIGRATION.md.
Finding: after seed + host, the self-hosted compiler uses no clojure.core fns
beyond the existing kernel tier (+ atom/swap!/reset! host primitives), so the
compiler-dep kernel tier is already complete — confirmed by self-host
conformance passing.
Add a maps-spec defspec covering nil values through the reader (literals), the
construction path, and the op surface (assoc/merge/merge-with/into/conj/zipmap/
select-keys/get-in/dissoc/reduce-kv, nil keys). Raise the clojure-test-suite
baseline 3919 -> 3926 to guard the assertions the fix unlocked.
With nil now preserved in maps, the ops that iterated via (keys (phm-to-struct
m)) or built struct/table results were still dropping nil-valued entries. Route
them through map-entries-of (all [k v] pairs incl nil-valued keys) and build
results via core-assoc / kvs->map / map-assoc1 (phm when a key/value is nil):
- merge, conj (map onto struct or phm), transient conj!
- merge-with: also use presence (contains?), not nil-of-value, to decide
whether to combine — a present nil triggers (f existing v), like Clojure
- reduce-kv: iterate phm-entries so nil-valued keys are visited
- select-keys / zipmap: collect entries, build via kvs->map
- get-in: walk with a sentinel so a present nil value isn't read as missing
conformance 218/218 x3, clojure.data/diff handles nil, suite green,
clojure-test-suite 3926 -> 3927.
Janet structs drop nil, so {:b nil}, (assoc {} :b nil) and friends silently
lost the entry — diverging from Clojure, where a nil value is still a present
key. Map construction now uses a phm (which preserves nil) whenever a key or
value is nil; the nil-free case stays a fast struct.
Touched the whole construction path: the reader (a nil-bearing map literal form
is a phm, not a struct), eval-form (evaluates phm map-literal forms, rebuilds
preserving nil), build-map-literal (compiled path), core-assoc (promotes on a
nil key/value), and the host map-form predicates h-map?/h-map-pairs (recognize
a phm form) so the self-hosted analyzer handles {:b nil} literals.
Safe now that the back end densifies phm IR nodes and the IR is nil-field-clean
(prior commits): const-nil nodes go phm but read back correctly, and the common
nodes stay structs, so fib stays compiled (0.08s, no interpret fallback).
Also fixes clojure.set/rename-keys, which deleted the old key via (assoc m old
nil) — a trick that only worked while nil was dropped; now uses dissoc.
conformance 218/218 x3, suite green, clojure-test-suite 3919 -> 3926.
def-node carried :meta nil when a def had no metadata, fn-node carried :name
nil for anonymous fns, analyze-arity carried :rest nil for fixed arities, and
empty-env carried :recur nil. These all mean genuine absence, and every
consumer reads an absent key back as nil — so writing the nil is gratuitous.
Dropping them keeps the common nodes (defs, fns, fixed arities) nil-free, so
once user maps preserve nil they stay plain structs instead of becoming phm.
Faithful nil values (a nil const :val) are left as-is; the back end handles
those. No-op today since structs already drop nil: suite, conformance 218/218
x3, fixpoint, fib all unchanged.
The back end reads IR node fields with raw (node :key) access, which assumes a
Janet struct. Once user maps preserve nil (coming), a node that carries a
nil-valued field — anonymous fn :name, nil const :val, def with no :meta, arity
with no :rest — is built as a phm instead, whose fields live under :buckets and
read back nil. norm-node densifies such a node via phm-to-struct, which drops
exactly those nil fields — precisely what the back end wants, since it already
treats an absent field as nil. Applied where a node first reaches the emitter
(emit entry, fn arities, invoke callee); structs pass through untouched.
No representation change yet, so this is a no-op today: full suite, conformance
218/218 x3, fixpoint, fib all unchanged. Keeps IR access a host-layer concern
so jolt.ir/jolt.analyzer stay portable.
phm-get returned the default for a key present with a nil value; fix it to
presence-check the bucket so a nil value reads back as nil, not the default.
core-keys/core-vals went through phm-to-struct, which drops nil-valued keys —
read phm-entries directly instead so those keys survive.
Isolated phm correctness only; no map-representation change. Battery holds at
3919, conformance 218/218 in all three modes. The earlier -8 came from
assoc's struct->phm promotion, not these accessors.
The back end emitted a direct Janet call (f args) for every :var/:local callee,
but Janet calling a pvec-table does get (no integer key -> nil), a keyword does
the wrong thing, etc. — so (let [v [1 2 3]] (v 0)), a keyword/meta-vector in a
binding, all returned nil in compile mode (interpret was fine). Now a direct call
is emitted only when the callee is provably a function: :fn / :host always, and a
:var whose current root is a function; everything else routes through jolt-call,
which dispatches IFn correctly (function fast-path first). User/core fn calls stay
direct, so no hot-path regression — fixpoint stage1==stage2==stage3 holds.
Conformance 218/218 all modes; full suite green. Spec cases added.
Document the two fixed behaviors in the contract spec: (with-meta x nil) clears
metadata, and a multi-collection map keeps nil elements (they're values, not
end-of-seq). The reader grammar (doc/grammar.ebnf) is unchanged — this session
touched semantics/stdlib, not surface syntax.
The reader yields set literals as a form ({:jolt/type :jolt/set ...}); edn now
walks the parsed result and builds actual sets (recursing into maps/vectors;
lists stay lists, never evaluated). Untagged-struct guard so symbols/chars/tagged
literals pass through. 47 -> 49 in the battery; only #uuid (no uuid type) remains.
clojure.edn/read-string now delegates to clojure.core/read-string (the
reader-based special form — parses, never evals) via a private helper, with the
opts-map :eof arity and nil/blank-input handling. 43 -> 47 in the stdlib battery.
This required an evaluator fix: special forms were matched by name only, so
clojure.edn/read-string dispatched the core read-string SPECIAL FORM instead of
the var. Now a head qualified to a non-core namespace resolves to its var; only
unqualified or clojure.core-qualified heads are special forms. Conformance
218/218 all modes; full suite green.
Remaining edn gaps (jolt-b7y): set literals read as a set FORM not a constructed
set, and #uuid has no real type.
The multi-coll map step detected end-of-seq with (nil? (ls-first cur)), which
also fired on a legitimate nil element — so mapping over a lazy-seq containing
nils truncated at the first nil. That's why (doall (map merge a b)) collapsed to
() as a reduce accumulator. Use seq-done? to detect exhaustion and push the value
(which may be nil). doall is unchanged; clojure.data restored to the canonical
(doall (map ...)) form. Conformance 218/218; battery 3919.
clojure.data: ported from the ClojureScript impl (jolt-native equality-partition
fn instead of host-type protocol dispatch). Verified against Clojure's own
data-test (12/13 canonical; the 13th needs nil-valued-map support, jolt-c7h).
Used mapv over the reference's (doall (map …)) — jolt's lazy multi-coll map +
doall don't force as a reduce accumulator (jolt-dzh).
clojure.zip: jolt's was a broken custom reimplementation; replaced with a port of
real clojure.zip (metadata-based locs). Uses (nth loc …) instead of (loc …)
because meta-bearing vectors aren't invocable as fns (jolt-vh5).
core: with-meta now accepts nil metadata (Clojure allows (with-meta x nil)); it
crashed calling keys on nil. This is what unblocked zip's make-node.
New vendored battery (test/clojure-stdlib/, from clojurust's suite with fixtures
corrected to match real Clojure): walk 34, zip 33, data 61 all clean; edn guarded
at 43 (still a stub — jolt-b7y). Conformance 218/218 all modes; full suite green.
The expensive multimethod path is the hierarchy fallback: when a dispatch value
isn't a direct method key, mm-fn walks every key with isa? (derive-based
dispatch). That resolution is now cached per dispatch value on the multimethod
var (:jolt/dispatch-cache), cleared in place by defmethod/prefer-method/
remove-method/remove-all-methods so a redefinition can't be served stale. Direct
(get methods dv) hits stay uncached — already a single lookup.
Test cases added to dispatch-cache-test (compile, interpret, aot-core off):
hierarchy hit is cached, a newly-added specific method is seen, removal re-exposes
the fallback. Conformance 218/218; full suite green.
Soundness gate for self-hosting (jolt-d0r). stage1 = analyzer built by the Janet
bootstrap; stage2 = analyzer rebuilt by compiling jolt.ir + jolt.analyzer through
stage1 (self-host); stage3 = the same self-rebuild again. A corpus of programs
must produce identical results across all three.
Compared BEHAVIORALLY rather than by emitted code: emitted Janet forms embed live
setter/getter closures (identity varies per compile) and the IR carries
representation-level gensyms and pvec internals, so structural equality fights the
representation. Behavioral parity is the property that actually matters and is
representation-independent.
Also corrects the battery baseline 3920 -> 3919: 3920 was a lucky run; 3919 is the
stable value (the 9 timeout-prone tests can shift the count by one).
