One rule, four positions, three treatments

.claude/skills/ci-gates/SKILL.md

@@ -8705,3 +8705,45 @@ The general form: a build graph is a claim about coverage, and unlike a
 test list nothing prints it. `import` is the edge, the root file is the
 whole specification of what gets compiled, and it is nine lines long.
 
+## 346. One rule, four positions, three treatments
+
+`types_compatible` rejects narrowing `F64 -> F32`, with a comment naming the
+issue it closed. Where does that rejection actually apply?
+
+| position | compared? | verdict |
+|---|---|---|
+| assignment `x = d;` | yes | **error** |
+| argument `p(d)` | yes | **warning** -- printed under a `Typecheck OK` header |
+| declaration `var x: f32 = d;` | **no** | silent |
+| return `-> f32 { return d; }` | **no** | silent |
+
+A soundness fix guarding one of four narrowing sites is a soundness fix in one
+of four narrowing sites. **When you find a rule, enumerate the positions it
+should hold in and check each one** -- the code will not tell you which ones it
+forgot, because forgetting is silent by construction.
+
+Adding the declaration comparison at the argument's severity -- a warning --
+cost 18 warnings in 13 files across the whole corpus, and moved no ratchet.
+I expected noise and got a work list.
+
+## 347. The check found the bug that made the check necessary
+
+Of those 18, two read `Str <- F64`:
+
+```
+var period_str : &str = "83.333";
+```
+
+`infer_expr` returns `Str` when the literal's VALUE starts with a quote. The
+parser marks the node `extra_kind: "string"` and the lexeme does not always
+carry the quote, so a quoted string fell through to the float branch and any
+string whose text parses as a number was typed as that number. `"hello"` was
+fine -- it does not parse as a float, so it landed on `Unknown`, which is
+compatible with everything and therefore silent.
+
+Two silences composed: the declaration position was never compared, and the
+value that would have failed the comparison was mistyped. Neither was visible
+alone.
+
+The fix reads the marker the parser already sets. `specs/pins/emitter_xdc.t27`
+now typechecks -- 627 to 628 specs, zero regressions.

.trinity/seals/EmitterXDC.json

@@ -1,6 +1,6 @@
 {
   "gen_hash_c": "sha256:038f5e24b6ee91e6b41023f5c1fa171c07d9ff4e1e8ab71e5ed6b0e0b26da53b",
-  "gen_hash_rust": "sha256:6b2272a70de5f6ff5633ddb7fb2ba8000f021bb2e4f4209b487770b620fca3a0",
+  "gen_hash_rust": "sha256:d307b19f6a47782c3a0ac22059dda6c8430f350f47956afd9bca952d57d91763",
   "gen_hash_verilog": "sha256:cde8b3da09478086d0bbe7f87fd950c2366b36c21e7bfb8d0bfc9718e995d1fb",
   "gen_hash_zig": "sha256:3f16baa3c8bd8fb3b2af6ceb5f59238e3646c841eec55979d286760c47211656",
   "module": "EmitterXDC",

.trinity/seals/pins_EmitterXDC.json

@@ -1,11 +1,11 @@
 {
   "gen_hash_c": "sha256:038f5e24b6ee91e6b41023f5c1fa171c07d9ff4e1e8ab71e5ed6b0e0b26da53b",
-  "gen_hash_rust": "sha256:6b2272a70de5f6ff5633ddb7fb2ba8000f021bb2e4f4209b487770b620fca3a0",
+  "gen_hash_rust": "sha256:d307b19f6a47782c3a0ac22059dda6c8430f350f47956afd9bca952d57d91763",
   "gen_hash_verilog": "sha256:cde8b3da09478086d0bbe7f87fd950c2366b36c21e7bfb8d0bfc9718e995d1fb",
   "gen_hash_zig": "sha256:3f16baa3c8bd8fb3b2af6ceb5f59238e3646c841eec55979d286760c47211656",
   "module": "EmitterXDC",
   "ring": 12,
-  "sealed_at": "2026-08-29T04:11:27Z",
+  "sealed_at": "2026-08-29T22:04:32Z",
   "sealed_by": "t27c-bootstrap@0.2.0",
   "spec_hash": "sha256:d9954ca190ee3e748ef8df7ce32bd183096f7f97003651aede1ccabe5c1b0e19",
   "spec_path": "specs/pins/emitter_xdc.t27"

bootstrap/src/compiler.rs

@@ -21897,6 +21897,35 @@ fn check_stmt(node: &Node, symbols: &mut Vec<SymbolEntry>, fns: &[FnEntry], resu
             for child in &node.children {
                 check_expr(child, symbols, fns, result);
             }
+            // #920 rejects narrowing F64 -> F32 in an ASSIGNMENT and nowhere
+            // else. Measured: `x = d` is an error, `p(d)` is a warning, and a
+            // DECLARATION with an annotation was not compared at all --
+            // `var x: f32 = d;` passed in silence. Three treatments of one
+            // rule, two of them silent.
+            //
+            // Reported at the same severity the argument position uses. Making
+            // it an error would be the consistent choice and would also fail
+            // specs that pass today; that is the owner's call, and a warning
+            // makes the hole countable without moving any ratchet.
+            if !node.extra_type.is_empty() && !node.children.is_empty() {
+                let declared = resolve_type_str(&node.extra_type);
+                let init = infer_expr(&node.children[0], symbols, fns);
+                if !types_compatible(&declared, &init)
+                    && declared != TypeInfo::Unknown
+                    && init != TypeInfo::Unknown
+                {
+                    result.warnings += 1;
+                    let line = if node.line > 0 {
+                        format!(":{}", node.line)
+                    } else {
+                        String::new()
+                    };
+                    result.errors.push(format!(
+                        "warning: '{}' is declared {:?} and initialised from {:?}{}",
+                        node.name, declared, init, line
+                    ));
+                }
+            }
             // #920 bug 2: register the local in the CURRENT scope so subsequent
             // sibling statements can resolve it (was previously discarded).
             symbols.push(SymbolEntry {
@@ -22110,7 +22139,14 @@ fn infer_expr(node: &Node, symbols: &[SymbolEntry], fns: &[FnEntry]) -> TypeInfo
             if node.value == "true" || node.value == "false" {
                 return TypeInfo::Bool;
             }
-            if node.value.starts_with('"') {
+            // The parser marks a string literal with `extra_kind: "string"`.
+            // This tested the VALUE for a leading quote instead, and the lexeme
+            // does not always carry one -- so `var period_str : &str = "83.333";`
+            // fell through to the float branch below and the declaration was
+            // typed F64. Measured across the corpus: two such strings, both in
+            // specs/pins/emitter_xdc.t27, and neither was reported by anything
+            // because the declaration position was not compared at all.
+            if node.extra_kind == "string" || node.value.starts_with('"') {
                 return TypeInfo::Str;
             }
             if node.value.parse::<i64>().is_ok() {

bootstrap/stage0/FROZEN_HASH

@@ -1 +1 @@
-f569da0c6c1415111f74b62680370272b7542ef1fe7f66c68267cbd71f2806df
+472b902a8cef8b3f673c72e8e2735bad6eff4511b6869e35a7ae178090e26b8c

docs/now/2026-08-30-one-rule-four-positions-three-treatments.md

@@ -0,0 +1,8 @@
+# NOW -- One rule, four positions, three treatments (2026-08-30)
+
+## One rule, four positions, three treatments (Refs #2864)
+
+- #920 rejects narrowing F64 -> F32. Measured where the rule actually applies: ASSIGNMENT is an error, an ARGUMENT is a warning printed under a 'Typecheck OK' header, and a DECLARATION with an annotation was not compared at all. A return value is not compared either.
+- Added the declaration comparison at the same severity the argument position uses -- a warning, so no ratchet can move. Cost across the corpus: 18 warnings in 13 files, not the noise I expected.
+- Two of those were the sharp ones: 'var period_str : &str = "83.333";' typed F64. infer_expr tests the VALUE for a leading quote while the parser marks the node extra_kind=string, and the lexeme does not always carry the quote -- so a quoted string whose text parses as a float became a float.
+- Fixed by reading the marker. specs/pins/emitter_xdc.t27 now typechecks: 627 -> 628 specs, zero regressions, bootstrap ratchet holds. 16 warnings remain -- 8 U64 <- F64, 7 F32 <- F64 -- countable debt rather than a silence.

tools/specs_generate_baseline.txt

@@ -83,4 +83,4 @@ specs/tri/collections/bitset.t27 | Error: Compile error: Expected LParen, got Kw
 specs/vm/jit_semantics.t27 | Error: Compile error: Expected LBrace, got Semicolon (';') at line 73:58
 specs/vsa/packed_vsa.t27 | Error: Compile error: Expected LBrace, got Semicolon (';') at line 58:74
 specs/vsa/sequence_hdc.t27 | Error: Compile error: Expected LBrace, got Semicolon (';') at line 96:88
-test_highlight.t27 | Error: Compile error: Unexpected top-level token: KwModule ('module') at line 4:5
+test_highlight.t27 | Error: Compile error: parse error at module level near line 21: Unexpected token in expression: Power ('**') at line 21:22