[yogthos]

"Epic jolt-ewxj. An audit measured the worst scorecard rows per operation inside one --opt binary and found five general mechanisms behind them, none of them the pass the row is named for. This lands the rounds that fix them, with the gates that pin each one.\n\nRuntime (round 1). A compare of two base scalars of different kinds, or of anything against nil, walked every registered eq arm before it was answered (145-240ns per miss; case lowers to that chain). It is answered ahead of the walk now, the cross-kind pairs joined the fast probes so the registry refuses an arm claiming one, and values-test asserts a registered arm is never consulted for them. jolt-wrap64/the unchecked helpers test fixnum? first (unchecked-add 12 -> 3ns), jolt->idx passes a fixnum through, str has 2/3-arg entries that render strings/keywords/fixnums without the printer (56 -> 19ns), pvec-nth! reads the tail without pv-leaf-for's two-value return, jolt-invoke builds its arity-error name only on the error path.\n\nEmitter (rounds 2-3, remint). The unchecked-* family is native (2-arg helpers directly; :lng/:dbl on the wrapping/flonum ops), unchecked-long/-int are casts typing their result :long, and the long casts pass a fixnum through inline: char-scan's loop counters are fixnums now. An interop call on an unproven receiver whose method has a string/keyword direct form tests the receiver at the site (receiver and args bound once, in order) with record-method-dispatch as the slow arm. Under direct-link a call to a seed var binds the root procedure once at load and calls it directly, where it paid var-cell-deref + jolt-invokeN per call; jolt.host/seed-callable? applies the closed-world rule (procedure root, arity admitted, not ^:dynamic/^:redef, not redefined by the app), same semantics as JVM direct linking, and jolt run never direct-links. A ^doubles param's backing flvector is bound once at the arity's entry (shadowing by let/loop/nested params/catch tracked; a non-array argument raises the checkcast ClassCastException on entry). case compares keyword/nil/boolean constants with identical?. JOLT_WP_TRACE=1 prints both fixpoints' iteration counts.\n\nRuntime (round 4). first answers a cell or a vector from the backing store with no seq cell (45 -> 12ns); host-table's sorted first and nio-file's dir-stream/path seq are registry arms instead of set!-wrappers in front of every seq in the program.\n\nRound 5, measured instead of built. With the new tracing, ring-app's whole-program fixpoint runs 5 priming + 2 main iterations + 1 field round: 7 walks of 3360 nodes, nowhere near the caps bead jolt-tg2s assumed, so a changed-defs worklist would save a few seconds of a ~35s build at the cost of the byte-identical acceptance criterion. Recorded on the bead; the lever is per-walk cost.\n\nFound along the way. (unchecked-long \\a) answered 97 where the JVM raises (caught by certifying the new corpus rows; fixed). test/chez/ffi-native-error-test.ss still passed ffi/write's arguments offset-first from before the argument-order change, writing a zero sixteen bytes past a four-byte buffer; the gate passed or died with SIGKILL in macOS's allocator depending on heap layout, and Guard Malloc named the overrun (fixed). bench/build-phases.sh counted the Chez runtime/app compile rows as jolt passes (fixed). examples/ring-app did not build (its time/db/crypto pins predated the :jolt/provides migration; jolt-lang/examples#35, merged).\n\nGates. New run-directcall.ss (seed direct-link, the interop guard, the unchecked lowering, case identity), flarr/traceemit follow the hoisted array shape, values-test gates the eq/first registries, ~100 corpus rows certified against JVM Clojure 1.12.5 (0 new divergences), make test green (95 CI targets, receipt on the pushed tree). make libconformance: 47 of 50 suites match their recorded tallies after the manifest fix; data.json, Selmer and ring-core differ from theirs only in a non-UTC zone (identical on a pristine-main binary, exact under TZ=UTC) — core's SimpleDateFormat renders UTC while jolt.time now applies the system zone, filed as jolt-6wm4.\n\nScorecard (same machine, same session, bench/run.sh before and after; opt ratio vs JVM Clojure):\n\n| row | before | after | jolt ms before → after |\n|---|---:|---:|---:|\n| char-scan | 24.2× | 6.7× | 363 → 94 |\n| literals | 7.5× | 2.1× | 195 → 53 |\n| string-ops | 6.9× | 3.7× | 511 → 250 |\n| arrays | 6.4× | 3.9× | 240 → 141 |\n| keyed-lookup | 5.7× | 3.5× | 147 → 95 |\n| string-build | 5.3× | 4.5× | 214 → 170 |\n| hash-eq | 3.0× | 2.6× | 537 → 467 |\n| transients | 3.8× | 3.5× | 222 → 201 |\n| seqs | 2.7× | 2.6× | 412 → 358 |\n| transducers | 4.2× | 4.2× | 139 → 128 |\n| sorted-access | 2.6× | 2.4× | 35 → 31 |\n| nth-access | 2.4× | 2.9× | 64 → 63 (the JVM column moved, jolt did not) |\n| mono-dispatch | 2.6× | 2.7× | 36 → 36 |\n\nThe ten rows without a same-session baseline (fib, tak, loop-recur, mandelbrot, mathfns, collections, vecops, dispatch, binary-trees, executors) were A/B'd against a pristine-main binary: every one within noise (branch/main 0.97-1.02×: fib 0.97, tak 0.97, loop-recur 1.00, mandelbrot 1.00, mathfns 1.00, collections 0.97, vecops 0.98, dispatch 1.02, binary-trees 0.98, executors 0.99). The full refreshed table is in bench/README.md.\n\nRound 6 (jolt-mw44.33): a small map is one flat slot vector. An array-mode map was a hash trie with an insertion-order list beside it — record, node, leaf pair per entry, order pair per entry — and every lookup hashed the key and descended the trie. It is now PersistentArrayMap's shape: a record over #(k0 v0 k1 v1 …) (chez-pmap-v5, image format 7; a chez-pmap-v4 in an old image re-mints through a legacy arm, and the v0.6.5 fixture exercises it). Lookup scans with a compare chosen once from the probe key's kind (keyword by identity, fixnum/string/symbol by their own equality, the rest by =); assoc/dissoc copy the slots; the transient is a slot buffer of max(16, len) slots with the reference's swap-on-remove; seq/keys/vals are vector-backed cells (O(1) count, chunk-loop reduce) instead of a cons chain built up front; reduce-kv is native and folds in place; assoc of the value already held and dissoc of an absent key answer the same map, as both JVM map classes do; into {} from a map and zipmap over vectors skip the seq entirely; two-entry dynamic literals build without a rest list.\n\nUnder it, a general fix: the runtime half of a binary compiles at Chez optimize-level 2 (level 3 is unsafe mode), so a checked scan of a 64-keyword map cost 3x the unchecked one and a 64-slot copy loop 390ns against 62. A new adapter tier (capability-unchecked: sa-ufx+, sa-uvector-ref …, Chez #3% forms, Gambit the checked primitives) carries the hot loops whose bounds are proven by construction, and every vector copy under a trie node update, tail append or transient growth is a bulk move. Two traps: irregex redefines vector-copy at top level as a one-argument loop, and on the Gambit host records are vectors, so the mode test is hnode?, never vector? (gambitboot caught it).\n\nSame-session probe, ns/op before → after: (:k m) 3-entry 34 → 12, 8-entry 25 → 15, (get m 20) 43 → 16, contains? 31 → 11, assoc replace 116 → 64, dissoc 111 → 59, {:a x :b y} at run time 100 → 17, (= m m') 230 → 84, (into {} m) 1045 → 205, zipmap 400 → 201, reduce-kv on 8 entries 583 → 50 and on 64 keywords 4259 → 324, {:keys [..]} destructure 96 → 48. Slower, and called out: the 64th key of a 64-keyword map 38 → 53 (the 30th 38 → 29; the scan is linear as PersistentArrayMap's is) and a symbol probed against symbol keys and missing 20 → 38 (a checked record predicate + accessor per slot at level 2; a symbol probed against keyword keys, honeysql's shape, is faster than before). Hash-mode maps are unchanged.\n\nCI. The red test job was one unit row: telemetry read jolt.host/bytes-allocated before and after building 100k strings and expected growth, but that reading is live data plus uncollected garbage, so a collection landing between the two reads makes it shrink (reproduced in bare Chez). The row now asserts the lower bound that holds whatever the collector does.\n\nGates. New make arraymap (representation per mode, thresholds, transient buffer, vector-backed views, identity), transient rows (200k zipmap linear, array-map source), 21 corpus rows certified against the JVM oracle (0 new divergences), unit rows for order/identity/lookup-by-kind/reduce-kv, adaptercheck/portcheck for the new tier, gambitboot.\n\nBench rows, A/B/A in one sitting against a worktree of the commit before (a jolt binary assembles a built app's runtime from the checkout's sources, so a "before" binary run inside this tree measures nothing — the first attempt came back flat on every row for exactly that reason):\n\n| row | before | after | before again | jolt/JVM now |\n|---|---:|---:|---:|---:|\n| keyed-lookup | 97 | 91 | 98 | 3.5× |\n| collections | 19.3 | 17.0 | 19.3 | 1.5× |\n| transients | 205 | 147 | 203 | 2.5× |\n| hash-eq | 461 | 443 | 458 | 2.4× |\n| literals | 53 | 54 | 53 | 2.3× |\n| seqs | 358 | 340 | 357 | 2.4× |\n\ncollections and transients are hash-mode maps: they move because every trie node copy and transient buffer growth is a bulk move now.\n"