fix(groovy): tri-comparison manual verification -- 3 real engine defects fixed

docs/language_status/README.md

@@ -70,7 +70,7 @@ epic #813), not that no cases exist.
 | embedded_python | production | line_exclusive | 51/52 | 64 | 107 | not written |
 | **[fortran](fortran.md)** | production | positional_anchored | 45/52 | 35 | 101 | **written** |
 | go | production | standard_block | 51/52 | 47 | 84 | not written |
-| groovy | production | standard_block | 44/48 | 53 | 91 | not written |
+| **[groovy](groovy.md)** | production | standard_block | 44/48 | 53 | 91 | **written** |
 | haskell | production | recursive_block_haskell | 52/52 | 48 | 97 | not written |
 | html | production | block_exclusive | 39/48 | 91 | 123 | not written |
 | **[java](java.md)** | production | standard_block | 50/52 | 70 | 91 | **written** |

docs/self_scan/manual_verification.json

@@ -48,6 +48,22 @@
         "verified_by": "Claude Sonnet 5"
       }
     },
+    "groovy": {
+      "class": {
+        "note": "class_start verified against the same 188-file language-crucible v1.2.0 groovy corpus, same independent grep-based ground-truth methodology as the function entry above (zero shared implementation, triple-quoted-string-aware). Fixed one real GitGalaxy defect found in the process: class_start's capture group was entirely missing (pattern.groups() == 0) -- class *existence* was detected correctly (struct_class_start's raw count matched real declarations exactly), but detector.py's _resolve_class_start_match could never extract a name from a group that didn't exist, so every single one of the 286 pre-fix class_data rows was literally named the generic placeholder 'Anonymous_Class' instead of its real name -- confirmed the only language in _CLASS_START_NAMED_EXTRACTION_LANGS with this defect (checked all 36). Post-fix: zero unexplained recall gaps found across the whole corpus, and a precision spot-check found no false-positive class of the kind found for functions (class_start requires an explicit class/interface/trait/enum/record keyword, a far more constrained grammar shape than func_start's zero-prefix branch, so the MarkupBuilder/DSL-call ambiguity filed as #2530 doesn't apply here). See docs/language_status/groovy.md \u00a79.",
+        "total": 222,
+        "verified": 222,
+        "verified_at": "2026-08-31",
+        "verified_by": "Claude Sonnet 5"
+      },
+      "function": {
+        "note": "func_start verified against the language-crucible v1.2.0 groovy corpus (188 files, 9 real-world repos -- fineract's Gradle plugin, Gradle's own build-logic/integTest sources, Spock's own core+smoke+spec suites, Jenkins .groovy view templates). tree-sitter has no usable grammar for this language (see NODE_MAPS's own groovy exclusion comment in tree_sitter_accuracy_audit.py) and ctags has zero Groovy support at all (confirmed via `ctags --list-languages`), so this is a genuine 1-tool language despite tree-sitter nominally loading a 'groovy' grammar. Verification method: an independent grep-based ground-truth scan (no shared implementation with func_start) across every one of the 188 files, cross-checked against the real pipeline's function_data output by name, with triple-quoted string spans excluded via quote-parity tracking (both sides must agree a candidate is real top-level code, not fixture text embedded in a multi-line string) -- zero unexplained recall gaps found after fixing three real GitGalaxy defects surfaced along the way: func_start's zero-prefix branch matched the bare 'def' keyword once a quoted Spock feature-method name got correctly blanked for brace-safety (fixed by excluding 'def' from the branch's own exclusion lookahead, plus preserving the quoted name in detector.py's brace-safety shield instead of blanking it); Prism never shielded groovy's \"\"\"/''' multi-line strings at all, letting embedded fixture/template code (Gradle build-script snippets, Spock's own compiler-smoke-test source samples) leak into the real code stream and corrupt extraction for the rest of the file; and _extract_name's generic last-token heuristic silently reduced a quoted multi-word Spock description down to its last word. A manual precision spot-check (50+ random named entries read against source) found one remaining, unfixed defect class: MarkupBuilder/Jenkins-Pipeline-DSL calls with a trailing closure (`button(name: \"clear\") { ... }`) are syntactically indistinguishable from a real zero-prefix declaration and get misdetected the same way -- filed as #2530 (a genuine grammar ambiguity needing its own design pass, not a drive-by regex fix). Confirmed confined to the jenkins_view_groovy/ corpus folder (29 files): all 57 functions detected there are this shape, subtracted from `verified` below. See docs/language_status/groovy.md \u00a79.",
+        "total": 718,
+        "verified": 661,
+        "verified_at": "2026-08-31",
+        "verified_by": "Claude Sonnet 5"
+      }
+    },
     "jcl": {
       "class": {
         "note": "class_start (JCL JOB cards) verified against the language-crucible v1.2.0 corpus (186 .jcl files, 6 repos; 185 pipeline-scanned, same DFH$SIP1.jcl zero-density exclusion as the function entry). Three checks over the whole corpus: (1) GitGalaxy's class_start regex vs. an independent JOB-card grep (^//[A-Za-z0-9@#$][A-Za-z0-9@#$]{0,7}[ \\t]+JOB([ \\t]|$), no shared implementation, independently rejects //*-commented job cards) -- exact agreement, 118 == 118, zero per-file mismatches. The 12 cics-banking-sample-application-cbsa/CREL0*.jcl + CREDB2L.jcl files whose entire JOB card is commented out (//*NAME JOB ...) and the 2 cics-genapp sampcma/sampwui.jcl angle-bracket placeholders (//<CMASAPPL> JOB) remain correctly NOT counted. (2) real pipeline DB output: struct_class_start == class_count == 118 on every file, class_data carries 118 rows (JCL is in _CLASS_START_NAMED_EXTRACTION_LANGS since #2203, no brace-slice fallback). (3) direct source read of the v1.2.0 delta: cics-java-jcics-samples/etc_VSAM_DEFVSAM.jcl contributes exactly one JOB card (//DEFVSAM JOB). Every corpus file carries at most one JOB card. Count moved 117 -> 118 on 2026-08-30 solely from that one new file (crucible v1.1.0 -> v1.2.0, gitgalaxy pin bump #2478). Zero false positives. See docs/language_status/jcl.md #9.",
@@ -73,7 +89,7 @@
         "verified_by": "Claude Sonnet 5"
       },
       "function": {
-        "note": "func_start named extraction, verified by a string-aware independent scanner + full name-diff against EVERY detected handler across the whole language-crucible/data/livecode corpus (98 ingested files): 964/964 handler headers map 1:1 to a real definition, ZERO false positives, ZERO missed. Both dialects handled after #2409/#2410: LiveCode Script `[public|private] on|command|function|getprop|setprop <name>` (unparenthesized args) and LiveCode Builder `.lcb` `[public|private] handler <name>(...)` (typed, parenthesized). `foreign handler` FFI binding declarations (C-prototype shaped, `binds to \"<builtin>\"`, no body) are correctly NOT counted, and `handler type <Name>()` (a function-pointer typedef) is not counted. `total` is GitGalaxy's own live func-claim count (964), the chart's staleness anchor (matches both func_recall.matched_consensus and func_precision.total_slots); this is a PRECISION record for handler *existence*. #2410 routed livecode to detector.py Mode D (keyword depth-tracking) -- before that, Mode B brace-slicing produced ONE FunctionNode for the entire corpus. Count moved 934 -> 964 on 2026-08-29 when #2419 was fixed: `_apply_literal_shield` / prism's `multi_style_live` string mask was C-escape-aware, so `\"\\\\\"` (a real one-char LiveCode string containing a backslash) was misread as an escaped quote, desyncing quote pairing until a `/*` inside a later string literal (`\"Android/*\"`) opened a bogus block comment that blanked ~30 real handlers in revsaveasandroidstandalone.livecodescript alone. Both the earlier 934 count AND its verification undercounted for the same reason. No truncation markers remain anywhere in the corpus. See docs/language_status/livecode.md §9.",
+        "note": "func_start named extraction, verified by a string-aware independent scanner + full name-diff against EVERY detected handler across the whole language-crucible/data/livecode corpus (98 ingested files): 964/964 handler headers map 1:1 to a real definition, ZERO false positives, ZERO missed. Both dialects handled after #2409/#2410: LiveCode Script `[public|private] on|command|function|getprop|setprop <name>` (unparenthesized args) and LiveCode Builder `.lcb` `[public|private] handler <name>(...)` (typed, parenthesized). `foreign handler` FFI binding declarations (C-prototype shaped, `binds to \"<builtin>\"`, no body) are correctly NOT counted, and `handler type <Name>()` (a function-pointer typedef) is not counted. `total` is GitGalaxy's own live func-claim count (964), the chart's staleness anchor (matches both func_recall.matched_consensus and func_precision.total_slots); this is a PRECISION record for handler *existence*. #2410 routed livecode to detector.py Mode D (keyword depth-tracking) -- before that, Mode B brace-slicing produced ONE FunctionNode for the entire corpus. Count moved 934 -> 964 on 2026-08-29 when #2419 was fixed: `_apply_literal_shield` / prism's `multi_style_live` string mask was C-escape-aware, so `\"\\\\\"` (a real one-char LiveCode string containing a backslash) was misread as an escaped quote, desyncing quote pairing until a `/*` inside a later string literal (`\"Android/*\"`) opened a bogus block comment that blanked ~30 real handlers in revsaveasandroidstandalone.livecodescript alone. Both the earlier 934 count AND its verification undercounted for the same reason. No truncation markers remain anywhere in the corpus. See docs/language_status/livecode.md \u00a79.",
         "total": 964,
         "verified": 964,
         "verified_at": "2026-08-29",