Perceus ARC backend (--memory=arc), osprey test line coverage, and standalone-file LSP diagnostics

[MelbourneDeveloper]

TLDR

Implements the Perceus ARC reference-counting backend (--memory=arc, plan 0011 phase 2) end-to-end — counting runtime, full compiler dup/drop insertion, and leak diagnostics — plus native line coverage for osprey test (--coverage / --coverage-json) surfaced in the VS Code Test Explorer, and LSP diagnostics for standalone module-bearing files.

Details

ARC memory backend ([GC-ARC-PERCEUS], spec 0018 [MEM-BACKENDS])

• New counting runtime compiler/runtime/memory_arc.c: 16-byte header (meta + refcount), probe-first open-addressing registry so foreign/rodata pointers are safe no-ops, kind/mask-driven drop walk (RAW / MASK / list-header variants), immortal values (rc < 0), worklist-based recursive drop, and OSPREY_ARC_DEBUG=1 leak statistics. memory_hooks.h + osp_arc_shim.h give all three backends (default / gc / arc) one IR: default and gc define osp_retain/osp_release/osp_mem_immortal as no-ops.
• Compiler dup/drop discipline in crates/osprey-codegen: new arc.rs ownership ledger inserts Perceus retain/release — producers own +1, dup-on-store, drops at region end and at last use (release_dead_after computes continuation free-idents via the new freevars::free_idents_of_stmts), returns transfer +1, callees borrow. Owned values spill to null-initialized entry-block slots, so untaken match arms/branches need no dominance analysis. New meta.rs computes the per-allocation kind/mask word for @osp_alloc_tagged (unit-tested against the Result, closure-cell, and HttpResponse layouts).
• Container/boundary soundness: list/map runtimes retain on alias returns, the empty-list singleton is immortal, elements are dup'd at insertion, osprey_map_merge releases unaliased intermediates, fibers release the spawned env cell after the thunk runs, and pointer-erasing boundaries (effect boxing, fiber thunks, fold accumulators) escape-retain before boxing. Container interior nodes stay RAW (leak-safe by construction — never a dangling release); node-level RC is the documented M4b/M5b/M6b remainder in plan 0011.
• --memory=arc is wired through the CLI/Makefile; make _conformance-arc runs the harness under the counting runtime.

Test coverage ([TESTING-COVERAGE-CLI]/[TESTING-COVERAGE-JSON], spec 0027)

• osprey test --coverage instruments each suite with line counters (new crates/osprey-codegen/src/coverage.rs + compiler/runtime/coverage_runtime.c), prints per-suite and total # coverage: TAP rows, and --coverage-json <path> writes the merged machine-readable report. Coverage never changes suite outcomes or exit codes.
• VS Code Test Explorer runs suites through --coverage-json and feeds line hits into the native VS Code coverage API (test-explorer.ts, test-explorer-parse.ts).

LSP

• diagnostics.rs gains standalone_diagnostics: module-bearing files no project claims (test suites, standalone scripts) are assembled the same way --check/osprey test does, so the editor stops reporting unknown identifier for a file's own modules (backed by the new osprey_project::needs_assembly).

Docs/examples: plan 0011 checklist updated to the landed state; spec 0018 status rewritten; Talon Bank gains accounts.test.ospml; committed wasm bundles regenerated for the new codegen.

How Do The Automated Tests Prove It Works?

• Backend conformance: the differential harness (crates/diff_examples.sh) passes PASS=148 FAIL=0 NOEXP=0 + FC_OK with byte-identical output under all three backends (default, --memory=gc, --memory=arc) — the [MEM-BACKENDS] conformance bar. With OSPREY_ARC_DEBUG=1, container-free examples (comprehensive_math, record_update_basic, algebraic_effects_comprehensive, fiber_determinism) report zero live language values at exit.
• Layout metadata: meta.rs unit tests (result_block_masks_depend_on_payload, http_response_marks_words_1_2_5, oversized_structs_fall_back_to_raw) pin the mask encoding to the real C ABIs, including the sign-bit word-55 edge.
• Runtime: coverage_runtime_tests.c covers the C coverage collector; the string/profiler C suites and cargo test --workspace (all crates, ~1100 tests) are green; per-crate coverage thresholds in coverage-thresholds.json all met ([rust] OK: all crates meet their thresholds, extension 97.7% ≥ 95%).
• Coverage feature: test_cmd.rs tests (only_test_suffixed_files_are_recognized, coverage_dump_path_is_stable_and_scratch_scoped, arg-parse tests for --coverage-json) plus extension suites test-explorer-parse.test.ts / test-explorer.test.ts (coverage args, JSON parsing, line-hit reporting) — extension suites 3/3 green under the real VS Code host.
• Cross-flavor: cross_flavor_ir_equiv proves .osp/.ospml twins still lower to byte-identical IR with ARC insertion enabled.
• End-to-end: bank native tests (osprey test, 13 TAP tests) and bank Playwright e2e (17 passed) run the compiled ARC-era binary serving real HTTP; wasm golden suite PASS=49 FAIL=0 proves the retagged allocator links and runs on wasm32-wasip1; website Playwright 76 passed; Docker web-compiler test passed.