wam_go: P3 uw-resolve (deb / Provides / alternatives)

docs/WAM_GO_STATUS.md

@@ -25,7 +25,7 @@ WAM pipeline via `prefer_wam(true)`.
 |---|---:|
 | `src/unifyweaver/targets/wam_go_target.pl` | ~4.3k |
 | `src/unifyweaver/targets/wam_go_lowered_emitter.pl` | ~0.8k |
-| Dedicated tests | ~23 files |
+| Dedicated tests | ~26 files |
 
 ## What's shipped
 
@@ -88,7 +88,7 @@ classes that are fleet-wide suspects. Go's audit:
 |---|---|---|---|
 | A1 | `sub_string/5` builtin missing | **verified missing** | Go has `sub_atom/5` (`state.go.mustache:2792`) but no `sub_string/5` |
 | A2 | Y-register clobber across `Call` of a no-`Allocate` fact | **aliasing form: verified (structural), Call Y-save is a partial mitigation, X101≡Y1 not hit by uw-resolve. Frameless-Y form: REPRODUCED as a live wrong answer on the lowered lane → FIXED 2026-09** | encoding `X_n→n+99 / Y_n→n+199`, so **X101 ≡ index 200 ≡ Y1**. Call snapshots Y 200..299 and Proceed restores; Execute does **not** push (LCO). Choice points snapshot the Y-save stack. The numeric alias is unchanged. See the frameless-Y section below. |
-| A3 | `Execute` of a builtin doesn't return to the continuation | **handled for known builtins; `call/1` now classified** | `BuiltinExecute` takes Proceed's return path **including** `popCallFrame`. Residual: a missed classifier entry still silently fails. `call/1` is `wam_go_direct_builtin` → `BuiltinCall`. `member/2` in uw-resolve is `BuiltinCall`. |
+| A3 | `Execute` of a builtin doesn't return to the continuation | **handled for known builtins; P3 helpers classified; unknown-builtin warn knob landed** | `BuiltinExecute` takes Proceed's return path **including** `popCallFrame`. Residual: a missed classifier entry still fails the query; `UW_WAM_WARN_UNKNOWN=1` now prints `[wam_go] call|execute of unresolved goal NAME failed` (off by default). `call/1`, `maplist/2-4`, `predsort/3`, `functor/3`, `arg/3`, `=../2` are `wam_go_direct_builtin`. Lowered `pred_to_go_call` also falls back to `executeBuiltin` then warn. |
 | A4 | String fidelity | **rung 0** | `value.go.mustache` has Integer/Float/Atom/Compound/Structure/List/Ref/Unbound — no string type; D37's double-quoted literals intern as atoms |
 
 Pattern lane: `go_target.pl` compiles facts from `clause(Head, true)`
@@ -234,16 +234,51 @@ runs `wpick` on the interpreter lane only and says why.
 
 ## Whole-program exercise: uw-resolve (`examples/pkg_resolver/go/`)
 
-P0.5 resolver compiled through `wam_go` (`prefer_wam(true)`). JSON shim
-is term↔JSON IO only. Corpus **39/39** vs SWI; seeded differential
-**2400/0**. Additional runtime bugs the program forced (not in the A2
-table): empty-list `GetConstant` vs `*List`; `sort/2` unique-collapsing
-compounds; `switch_on_structure` emission using `Val` instead of
-`Functor`; 4-arg `begin_aggregate`; `allocVarId` aliasing driver Idx
-10000–10999 (B3 `sort/2` unifying Acc with `[]`). **B3** on the 5k
-catalog (`0xc0ffee01`): Go load **0.060s** / resolve **4.604s**, same
-10-package selection as SWI (load **0.428s** / resolve **0.008s**). See
-`examples/pkg_resolver/go/README.md`.
+P3 resolver compiled through `wam_go` (`prefer_wam(true)`). JSON shim
+is term↔JSON IO only (deb versions, Provides 3-/4-ary, alternatives
+groups, `catalog/6|9|10`, blocked `providers`/`alternatives` shapes).
+Corpus **51/51** vs SWI; seeded differential **2600/0** (2400 `g*` @
+`0xa5b6c7d8` + 200 `p3g*` @ `0xdeb00001`). D59 `index_threshold(64)`
+comes free with the current `resolver.pl`.
+
+P3 builtin inventory (generated WAM → runtime → `wam_go_direct_builtin`):
+
+| Name | Found | Implemented | Classified |
+|---|---|---|---|
+| `maplist/2,3,4` | missing; compiler emits `BuiltinCall`/`Call` | yes, meta-calls user preds via `invokeGoalOnce` | yes |
+| `predsort/3` | missing; resolver `cmp_ver/3` for `deb/3` | yes | yes |
+| `functor/3`, `arg/3`, `=../2` | implemented, unclassified (A3) | already in `executeBuiltin` | now classified |
+| `string_codes/2` | present; empty result was `*List{}` | empty goes through `listFromItems` | already classified |
+| `memberchk/2` | already OK | — | — |
+
+`UW_WAM_WARN_UNKNOWN=1` (off by default) prints
+`[wam_go] call|execute of unresolved goal NAME failed` after the
+label/foreign/fact miss — additive, mirrors Rust.
+
+Bugs this program forced beyond the P0.5 table:
+
+| Symptom | Cause | Fix | Probe |
+|---|---|---|---|
+| `provides_virtual_only` / first-listed / versioned-ok → `{fail:true}` | `pick_need` indexes on Mode; `classic` is switch `"default"` with two clauses; Pc rewrite jumped to `idx+1` then `indexedClauseBodyStart` skipped `TryMeElse` | label-form default falls through (`PC++`); Pc form keeps the jump when the target is a try/retry head | `tests/test_wam_go_switch_default_chain.pl` |
+| 40 `p3g*` differentials `{fail:true}` (corpus still 51/51) | `predsort` meta-called `cmp_ver`; `invokeAtPC` smashed A2; `Unify(getReg(2), Sorted)` compared a version to a list. Single-element lists and `compare/3` never ran the less-func — corpus P3 rows filtered to one candidate | capture A2 before the comparator; restore regs after each `invokeGoalOnce`; leftover CPs truncated | `tests/test_wam_go_maplist_predsort.pl` (`gpreduser`) |
+
+**B1–B3** (this Cloud Agent VM, SWI 9.0.4, Go 1.22.2, Node v22):
+
+| Bench | Number |
+|---|---|
+| B1 corpus wall | **0.053s, 51/51** |
+| B2 differential | **2600 cases, 0 divergences**; SWI **1.671s** / Go **17.179s** |
+| B3 5k `resolve_layered` (`0xc0ffee01`) | load **0.065s** / resolve **10.620s**, same 10-package selection as SWI / D51 |
+
+Index effect: the 5k catalog is far above `index_threshold(64)`, so
+`resolve_layered` wraps `icat/3` (cannot toggle the frozen fact from
+owned files). D51 pre-index Go B3 was **11.0s** linear scans; with-index
+Go is **10.62s** (~1.04×). wamjs saw **9.2×** on the same G1 trees —
+Go's interpreted index build (~45 WAM instructions/row × 5k) almost
+cancels the lookup win. Selection matches the historical SWI 10-package
+set, so the index is semantically active.
+
+See `examples/pkg_resolver/go/README.md`.
 
 ## Path forward
 
@@ -278,3 +313,9 @@ now live on `ChoicePoint.Levels`, never in `vm.Regs`. Probe
 lowered-lane defect (a lowered method that `call`s the interpreter and
 then tail-recurses loses its output binding) is recorded as a residual
 above — it is not the barrier and is not fixed here.
+2026-09-04: **uw-resolve P3 on Go.** maplist/predsort/functor classified;
+`UW_WAM_WARN_UNKNOWN`; shim speaks deb/provides/alternatives; switch
+default try/retry chain; predsort A2 smash. Corpus **51/51**,
+differential **2600/0**, cut **35/35**, full `test_wam_go_*.pl` sweep
+green, shared-lane JS/cli_args/wamjs intact. B3 with-index **10.62s**
+vs D51 pre-index **11.0s**.

examples/pkg_resolver/go/README.md

@@ -1,10 +1,13 @@
 # uw-resolve on Go (`wam_go`)
 
-Compile the frozen P0.5 package resolver (`../resolver.pl`) through
+Compile the frozen P3 package resolver (`../resolver.pl`) through
 the Go WAM backend and drive it with a JSON shim. The shim converts
-catalogs/requests/env **in** and selections/explanations **out**. It
-contains no resolver logic (no candidate order, no constraint
-arithmetic, no layer walk).
+catalogs/requests/env **in** and selections/explanations **out**,
+including `{"deb":[Epoch,[[order,num],…],[…]]}`,
+`{"alternatives":[…]}`, provides 3- and 4-ary rows, `catalog/10`,
+and `blocked(..., providers([...]))` /
+`blocked(alternatives([...]))`. It contains no resolver logic (no
+candidate order, no constraint arithmetic, no layer walk).
 
 ## Build / run
 
@@ -42,6 +45,8 @@ times `resolve_layered` on `DIR/rich.jsonl` + `DIR/probe.json`.
 | 7 | bagof/setof probes left `L` as `__R200` | 4-arg `begin_aggregate bagof, Y, X, ''` was dropped as `// TODO`, so `EndAggregate` was a no-op | Accept the 4th witness-register argument (empty witness list uses the 3-arg runtime) | cut-semantics p14/p15 |
 | 8 | Cut suite lost extra solutions / p29 printed `1` not `ok` | Y slots are global; Proceed popped a Call Y-save that backtrack did not restore; `!/0` used the caller's `EnvFrame.CutB0` | Call pushes B0+Y; Execute rebases B0 only (LCO); Proceed pops; CPs snapshot/restore both stacks; `!/0` truncates to `PendingB0` | `tests/test_wam_go_cut_semantics.pl` (35/35) |
 | 9 | B3 `resolve_layered` unified Acc with `[]` | `allocVarId` started at 1000; shim minted output vars at `Idx: 10000+i`; after ~9000 cells, `NextVarId` aliased `Bindings[10002]` with `Selection` | skip 10000–10999 in `allocVarId`; shim uses `vm.allocVarId()` | `tests/test_wam_go_varid_collision.pl` |
+| 10 | P3 `provides_*` corpus rows `{fail:true}` | `SwitchOnConstantPc` default resolved to the `TryMeElse` then skipped it, so only the first classic clause of `pick_need/8` ran | fall through label-form default; keep Pc jump when the target is a try/retry head | `tests/test_wam_go_switch_default_chain.pl` |
+| 11 | 40 generated `p3g*` cases `{fail:true}` | `predsort(cmp_ver)` smashed A2; Unify then compared a version to the sorted list. Corpus survived because those P3 rows had one matching version (less-func never ran) | capture A2; restore regs after each comparator; `invokeGoalOnce` drops leftover CPs | `tests/test_wam_go_maplist_predsort.pl` |
 
 ## Cut / choice-point barriers
 
@@ -68,21 +73,25 @@ that path.
 
 ## Benchmarks (B1–B3)
 
-Measured on this Cloud Agent VM (SWI 9.0.4, Go 1.22.2). Dump corpus
-has 39 cases (plunit `test_resolver.pl` is 38 tests; the extra dump
-row is `alias_request_edge`).
+Measured on this Cloud Agent VM (SWI 9.0.4, Go 1.22.2, Node v22). Dump
+corpus has **51** cases (40 pre-P3 + 11 P3). Differential is 2400 `g*`
+(seed `0xa5b6c7d8`) + 200 `p3g*` (seed `0xdeb00001`).
 
 | Bench | What | Number |
 |-------|------|--------|
-| B1 | 39-scenario corpus, Go binary, one process (`run_corpus_go.sh`) | **0.024s wall, 39/39 vs SWI** |
-| B1 | `go build` of `uwresolve` (not in B1 wall; clean rebuild) | **0.168s** |
-| B2 | Differential Go leg (`run_differential_go.sh`, seed `0xa5b6c7d8`) | **8.869s, 2400 cases, 0 divergences, 0 crashes** |
-| B2 | SWI oracle leg, same machine | **1.331s** |
-| B3 | 5k catalog `resolve_layered` (`run_scale_go.sh`, seed `0xc0ffee01`) | Go **load 0.060s / resolve 4.604s**, 10-package selection `p1-v(1,0,0)`, `p10`, `p12`, `p2`, `p3`, `p30`, `p4`, `p5`, `p6`, `p7` |
+| B1 | 51-scenario corpus, Go binary (`run_corpus_go.sh`) | **0.053s wall, 51/51 vs SWI** |
+| B2 | Differential Go leg (`run_differential_go.sh`) | **17.179s, 2600 cases, 0 divergences, 0 crashes** |
+| B2 | SWI oracle leg, same machine | **1.671s** |
+| B3 | 5k catalog `resolve_layered` (`run_scale_go.sh`, seed `0xc0ffee01`) | Go **load 0.065s / resolve 10.620s**, 10-package selection `p1-v(1,0,0)`, `p10`, `p12`, `p2`, `p3`, `p30`, `p4`, `p5`, `p6`, `p7` |
 
-SWI term reference on the **same** `go/.scale_out` catalog + `p30` probe
-(not `run_scale_demo.sh`, which also builds the store path): **load
-0.428s / resolve 0.008s**, same 10 packages.
+Index: `index_threshold(64)` is a frozen fact; the 5k catalog wraps
+`icat/3`. D51 pre-index Go B3 was **11.0s**; with-index is **10.62s**
+(~1.04×). wamjs saw 9.2× on the same trees — Go's interpreted index
+build cost nearly cancels the lookup win. Threshold itself cannot be
+toggled from this directory.
+
+`loadRichJSONL` keeps provides rows and alternatives `dep` objects (the
+P3 overlay the older scale path stripped).
 
 ## Residuals
 
@@ -91,8 +100,10 @@ SWI term reference on the **same** `go/.scale_out` catalog + `p30` probe
   for Allocate-less Y clobber; the numeric alias is unchanged. The P0.5
   resolver did not hit a no-`Allocate` fact with >99 X placeholders.
 - **A3**: `BuiltinExecute` covers known builtins. A builtin missed by
-  translation-time classification still fails silently. `call/1` is
-  now classified. `member/2` is `BuiltinCall` here.
+  translation-time classification still fails the query;
+  `UW_WAM_WARN_UNKNOWN=1` prints the miss (off by default). `call/1`,
+  `maplist/2-4`, `predsort/3`, `functor/3`, `arg/3`, `=../2` are
+  classified. `member/2` is `BuiltinCall` here.
 - **call/1 extra solutions**: nested user goals are first-solution;
   opaque `!` is correct.
 - **bagof witness grouping** is not implemented (empty witness list is
@@ -101,7 +112,7 @@ SWI term reference on the **same** `go/.scale_out` catalog + `p30` probe
   general heap allocator. Programs that allocate >9000 unbound cells
   still wrap past the window; the shim now uses `vm.allocVarId()` so
   output vars never sit in that hole.
-- P0.5 resolver residuals (unchanged, frozen source): Provides/virtual,
-  Debian epoch/tilde, write paths, incremental stores, pkg CLI
-  (concurrent), per-file SFS layers.
-- Do not create `cljs/` or `rust/` — concurrent rounds.
+- Frozen `resolver.pl` residuals: Breaks≡Conflicts, Pre-Depends
+  ordering, Recommends/Suggests, multiarch, epoch display `0`, write
+  paths, incremental stores, pkg CLI, per-file SFS layers.
+- Do not edit anything under `examples/pkg_resolver/` except `go/`.

examples/pkg_resolver/go/atoms.go

@@ -9,20 +9,28 @@ package wam
 // Interned atom literals (compile-time deduplicated)
 var (
     wamAtom____0 = internAtom("[]")
-    wamAtom_blanket_1 = internAtom("blanket")
-    wamAtom_abi_anchor_2 = internAtom("abi_anchor")
-    wamAtom_over_frozen_3 = internAtom("over_frozen")
-    wamAtom_none_4 = internAtom("none")
-    wamAtom_classic_5 = internAtom("classic")
-    wamAtom_layered_6 = internAtom("layered")
-    wamAtom_from_catalog_7 = internAtom("from_catalog")
-    wamAtom_from_base_8 = internAtom("from_base")
-    wamAtom_any_9 = internAtom("any")
-    wamAtom_no_candidate_10 = internAtom("no_candidate")
-    wamAtom_modified_11 = internAtom("modified")
-    wamAtom_footprint_12 = internAtom("footprint")
-    wamAtom_layer_shadow_13 = internAtom("layer_shadow")
-    wamAtom_base_14 = internAtom("base")
+    wamAtom_unsatisfiable_1 = internAtom("unsatisfiable")
+    wamAtom_blanket_2 = internAtom("blanket")
+    wamAtom_abi_anchor_3 = internAtom("abi_anchor")
+    wamAtom_over_frozen_4 = internAtom("over_frozen")
+    wamAtom_t_5 = internAtom("t")
+    wamAtom_none_6 = internAtom("none")
+    wamAtom___7 = internAtom("<")
+    wamAtom___8 = internAtom(">")
+    wamAtom___9 = internAtom("=")
+    wamAtom_any_10 = internAtom("any")
+    wamAtom_layered_11 = internAtom("layered")
+    wamAtom_classic_12 = internAtom("classic")
+    wamAtom_from_catalog_13 = internAtom("from_catalog")
+    wamAtom_from_base_14 = internAtom("from_base")
+    wamAtom_unversioned_15 = internAtom("unversioned")
+    wamAtom_no_candidate_16 = internAtom("no_candidate")
+    wamAtom_modified_17 = internAtom("modified")
+    wamAtom_footprint_18 = internAtom("footprint")
+    wamAtom_layer_shadow_19 = internAtom("layer_shadow")
+    wamAtom_base_20 = internAtom("base")
+    wamAtom_is_v3_21 = internAtom("is_v3")
+    wamAtom_cmp_ver_22 = internAtom("cmp_ver")
 )
 
 // atomInternMap is the single source of pointer identity for atoms.