FEATURE-ci-darwin-focused-diagnostics: limit inferior loader override to package libraries

Remove the diagnostic-only LLVM libc++ DYLD directory while preserving actual build/prefix and provider library paths, clean debugger environment, selected compiler/rpath metadata and source instrumentation.

Existing real subprocess boundary probe: two expected RED failures, then all six tests pass. Actionlint, five shell syntax checks, embedded Python syntax and diff checks pass. Native test bodies remain unobserved; caller delta review and corrected native continuation remain pending.

Author: etr

.github/workflows/release.yml

@@ -195,7 +195,7 @@ jobs:
                           PKG_CONFIG_PATH='', CXXFLAGS='-O0 -g -DLHS_DARWIN_FOCUSED_DIAGNOSTICS')
           provider = root / 'prerequisites/provider'
           base_env['LDFLAGS'] = '-Wl,-rpath,' + str(provider / 'lib')
-          inferior_library_path = os.pathsep.join([str(provider / 'lib'), os.environ['V3_PACKAGE_TOOLCHAIN'] + '/lib/c++'])
+          inferior_library_path = str(provider / 'lib')
           try:
               save()
               if platform.system() != 'Darwin' or os.geteuid() != 0:

specs/tasks/FEATURE-ci-darwin-focused-diagnostics.md

@@ -1,6 +1,6 @@
 # FEATURE-ci-darwin-focused-diagnostics
 
-Status: In Progress — first native attempt failed in launcher setup before all six programs; local environment correction passed focused checks and awaits caller review.
+Status: In Progress — corrected native attempt built all six programs, but every launch stopped in dyld before a test body; the second loader setup correction passed local checks and awaits caller delta review.
 
 Base: local `v3` at `a5a72356741300e5da16b94090f0f26a67a2d1f7`.
 Branch: `task/FEATURE-ci-darwin-focused-diagnostics`.
@@ -15,12 +15,15 @@ This is temporary diagnostic instrumentation. This branch, including its tempora
 - [x] Prepare manual-only, read-only native Darwin batch under existing default-branch dispatch-capable `release.yml` filename. The required `tag` is a fixed diagnostic sentinel; it never selects a ref, creates a tag or publishes anything.
 - [x] Compile syntax, diagnostic logger smoke checks, workflow/embedded Python/shell syntax, actionlint, changed-file cpplint and diff checks locally; no native tests or full suites in this phase.
 - [x] Caller independently validates the diagnostic diff and reviews the workflow before publishing this branch only.
-- [x] Caller dispatched the first attempt and retained its source-bound failed setup receipt and raw logs; all six selected programs remain not-executed.
+- [x] Caller dispatched the first attempt and retained its source-bound failed setup receipt and raw logs; that receipt records all six selected programs as not-executed.
 - [x] Correct the workflow host/debugger versus inferior environment boundary and verify it with real local subprocess checks, syntax and actionlint.
-- [ ] Caller reviews the narrow setup delta before publishing the correction and continuing the same authorized six-program batch.
+- [x] Caller reviewed the first setup delta, published the correction and dispatched the corrected continuation.
+- [x] Retain the corrected native receipt: both library builds and all six program compilations passed; six single launch attempts stopped before any test body in dyld.
+- [x] Remove the diagnostic's additional LLVM loader override while retaining actual build-or-prefix/provider library directories, host isolation and selected compiler/linkage metadata; verify the real local subprocess boundary.
+- [ ] Caller reviews the second narrow setup delta before publishing the correction and continuing the same authorized six-program batch.
 - [ ] Caller records actual native results, context differences and causal assessment without reopening closed probes.
 
-Corrected continuation after caller delta review/publication (no selected program execution budget was consumed):
+Corrected continuation after caller delta review/publication (six launch attempts have occurred; zero test bodies have been observed):
 
 ```bash
 gh workflow run release.yml --ref task/FEATURE-ci-darwin-focused-diagnostics -f tag=darwin-focused-once-a5a72356
@@ -46,13 +49,44 @@ This sealed validation is historical diagnostic readiness; it does not establish
 
 First native attempt `40052229400`, job `120747190075`, source `8f9722656397961dc9b5d1af360552541a6ff398`, failed before bootstrap/build: `xcrun lldb --version` exited `-6` under the workflow's global provider/Homebrew LLVM `DYLD_LIBRARY_PATH`. The retained log reports missing Apple libc++abi symbols, a pure virtual call and mutex fatal error. Artifact `12544022476` SHA-256 `269c9db7bb48f58be881c5aa700cfc74c73605b48969d1334730f3fae82ea31d` was verified; its original receipt remains unchanged at the journal's `first-native-evidence/receipt.json`. All six cases are `not-executed`; no C++ result or cause was established.
 
-The correction removes inferior-only loader/tracing variables from host compiler/build/inspection/debugger environments. It retains the existing selected compiler wrapper and provider linker rpath, and passes the complete build-or-installed-prefix, provider and selected toolchain library path explicitly through LLDB `target.env-vars`. Header/storage tracing stays in the relevant inferior environment. Each case records that target environment separately from the debugger command environment. No C++ instrumentation, test assertion, timeout or runtime decision changed.
+The correction removes inferior-only loader/tracing variables from host compiler/build/inspection/debugger environments. It retains the existing selected compiler wrapper and provider linker rpath, and passes the actual build-or-installed-prefix and provider library directories explicitly through LLDB `target.env-vars`. The selected C++ library remains bound by the compiler wrapper's existing absolute linkage/rpath, with toolchain/SDK metadata and actual `otool` evidence retained separately. Header/storage tracing stays in the relevant inferior environment. Each case records that target environment separately from the debugger command environment. No C++ instrumentation, test assertion, timeout or runtime decision changed.
 
 Focused proof lives in the primary journal's `setup-repair/test_environment_boundary.py`: the original workflow failed the boundary assertions; all six tests pass against the correction. Tests execute the actual embedded workflow environment construction, `run_case`, generated LLDB settings, inspection call sites and command logging/monitoring with real local Python child processes. Only unavailable `xcrun`/inspection executable argv are adapted to the local environment reporter; there is no simulated LLDB result, Darwin execution, selected program launch or native success claim. They verify clean host/debugger environments, both complete inferior library paths, trace isolation, selected compiler/provider flags and no mutation back into build/metadata environments. All five shell steps and both embedded Python scripts parse; actionlint 1.7.7 and diff whitespace checks pass. Native setup/compile/execution remain unverified locally and require the caller's corrected native continuation after delta review. No full suites, native execution, publication, dispatch, merge or cleanup were performed during this repair.
 
 Narrow setup delta validation `8e28dc133b84fcaf143113acd67473a5` sealed PASS
 for `61a4647da4a85a4bf74759bc1cb11b7d828bf557`: one iteration, zero fixes,
 zero new unworked findings; eight fresh affected-domain approvals and explicit
 unchanged performance-invariant carry. Original sealed review/report preserved.
-Native setup and all six executions remain incomplete pending corrected dispatch.
+At that seal, native setup and all six executions remained incomplete pending corrected dispatch.
 The first workflow attempt ran zero selected programs.
+
+Corrected native attempt `40053955397`, job `120752645361`, source
+`2d6233d0d892d3e41f165c31851efcab8cc125ad`, passed host LLDB inspection,
+bootstrap, both TLS library builds and all six program compilations/links. Each
+program launched once and stopped in dyld before its first test body, reporting
+missing typed-allocation weak exports `__ZnwmSt19__type_descriptor_t` /
+`__ZnamSt19__type_descriptor_t`. No original assertion or scalar trace was
+observed; the six original failure causes remain unknown. Artifact `12545170653`
+SHA-256 `6b9bc155e41bac354446bcfc27399cb1ffe84f43717fb7a65c3a06c0e43d3507`
+was verified, and the original failed receipt and six selected logs remain
+unchanged in the journal's `corrected-native-evidence/` directory.
+
+The second setup correction removes only the extra LLVM `lib/c++` entry from
+the diagnostic inferior's `DYLD_LIBRARY_PATH`. The owning package runtime policy
+limits loader overrides to package/provider/prerequisite directories and keeps
+toolchain identity separately; this diagnostic retains its existing build-or-prefix
+`lib` and provider `lib` paths rather than reproducing every package `lib`/`bin`
+entry. All six actual `otool` logs already identify the selected libc++ absolute
+path. This is a required setup correction, not a proven C++ cause or repair.
+
+Updated existing task-local boundary tests failed exactly the two inferior search
+expectations against the prior workflow (`setup-repair/second-red.log`), then all
+six passed against the one-line workflow correction (`second-green.log`). They
+exercise real workflow subprocess environment selection and generated LLDB target
+settings, preserving build/prefix/provider paths, trace isolation, host isolation
+and compiler/SDK/provider/toolchain metadata. Native tools' argv alone is adapted
+to a local environment reporter; no native binary or simulated Darwin result is
+executed. Shell and embedded Python syntax, actionlint and diff checks pass.
+The original historical validations and unread helper reports remain untouched.
+The second correction requires fresh caller delta review before publication or
+native continuation; no native check has been rerun and its gates remain incomplete.