Changelog

​

Improved

• Parser robustness. The disassembler now handles the constructs real Emscripten modules emit without desyncing: exception-handling opcodes (try, catch, throw, rethrow, delegate, catch_all, try_table), tail calls (return_call, return_call_indirect), and typed block types. Unknown sections (data count, tag, custom) and unmodeled opcodes are consumed and skipped rather than crashing.
• Float decompilation. The lifter renders f32 and f64 arithmetic, comparisons, and constants as readable expressions, and degrades SIMD (v128) opcodes to a comment.
• Oracle calibration. Deterministic identify tie-breaking that does not depend on store order, calibrated infer_version confidence, SignatureStore.stats() with the warden oracle inspect command, and an evaluate_identification precision/recall harness for measuring a corpus without emsdk.
• Sharper similarity. The fuzzy score blends the MinHash term with call-neighborhood, opcode-histogram, type-signature, and instruction-count terms and exposes its sub-scores, and the diff requires a shared structural skeleton before a fuzzy modified pairing. Stable identity and determinism are unchanged.
• Per-role specialists. A SpecialistRouter routes the naming and analyzer roles to separate configured backends or models. The default is one auto-selected backend, unchanged and offline-first.
• Float execution. The interpreter executes f32 functions, and differential_versions generates deterministic float inputs and compares them NaN-safely, so behavioral equivalence now covers float functions with no native toolchain.
• Richer dashboard. The read-only dashboard gained symbol search (GET /api/search), a function detail panel with evidence, and a time-travel history panel (GET /api/history/{stable_id}), all under the same query_only read-only guarantee.

​

Added

• Oracle corpus farm. harvest_directory(root) and warden oracle harvest <dir> build one oracle.json from a built matrix directory and its manifest.json. build_corpus_from_manifest reads a manifest of labeled modules. classify_library now spans musl, musl-pthread, libc++, libc++abi, compiler-rt, dlmalloc, emscripten, and wasi-libc. SignatureStore.extend deduplicates so re-runs stay idempotent, and SignatureIndex.build accepts bands=None to auto-pick the band count via recommended_bands. build_matrix.sh now compiles .cpp with em++ and writes a manifest, with new pthread, allocation, and C++ reference programs.
• Time-travel queries. KnowledgeBase.when_first_seen, evolution_of (with a body_changed flag), symbol_history, find_by_name, and resolve_stable_id. The warden history <name-or-id> command shows when a function first appeared, how it evolved, and who named it. diff_versions gained a store=False flag for pure-read diffs.
• Specialist crew and MCP tools. The concurrency and struct analyzers run inside the call-graph crew as economy-gated proposers, and run_agent_pass(..., rounds=N) walks the call graph bottom-up to a fixpoint. New MCP tools: search_symbols, get_diff, export_kb_text, and analyze_version.
• Behavioral verification. verify.corpus.generate_inputs builds a deterministic input corpus. differential_versions proves two versions of a function behave the same using the interpreter, with no native toolchain; warden equiv <from> <to> runs it. The interpreter now models signed and unsigned division and remainder, shifts and rotates, bit-counting, unsigned comparisons, select, and narrow 8 and 16 bit loads and stores, and traps cleanly on division by zero. run_differential orchestrates the wasm2c/w2c2 lift, compile, and compare pipeline through an injectable runner, and reports an honest plan when the C toolchain is absent.
• Interactive UX. A new warden.ui package: a rich terminal diff view (warden ui diff <from> <to>) and a read-only HTTP dashboard (warden serve) built on the standard library, with a confidence heatmap, coverage summary, and from/to diff over a small /api/* JSON surface. The connection runs in query_only mode, so the HTTP surface cannot write.

​

Added

• Ingestion. Pure-Python WebAssembly binary parser covering LEB128, all standard sections, and full opcode disassembly including sign-ext, bulk-memory, reference types, threads/atomics, and SIMD immediates. Includes the name custom-section parser and an Emscripten JS-glue parser (version, dynCall signatures, pthread/PROXY_TO_PTHREAD markers).
• Knowledge base. SQLite-backed, versioned symbol store keyed to stable function identities, with provenance, confidence, evidence, struct layouts, a thread/memory model, an audit log, and the provenance/confidence economy enforced at the write layer.
• Stable identity and fingerprinting. Structural skeleton hash, opcode-class histogram, call-neighborhood, surviving type signature, and a deterministic MinHash fuzzy signature. One composite similarity engine reused by both the Oracle and the diff.
• Emscripten Oracle. Signature store, corpus builder from labeled modules, identification pass, and Emscripten-version inference. The warden.oracle.index module adds SignatureIndex.build(store, *, bands=8), index.candidates(fp), and identify_indexed(kb, version_id, store, *, threshold=0.82, write=True), a band-based MinHash-LSH structure for sublinear candidate lookup that matches the linear identify() at the same threshold. CLI: pass --indexed to warden oracle identify. Containerized emsdk matrix scaffold under scripts/corpus/.
• Cross-version diff and carry-over. Match/classify pipeline (unchanged / moved / modified / new / deleted), automatic annotation carry-over (verbatim for shared identities, penalized for fuzzy matches), and a semantic changelog separating app changes from runtime churn.
• Built-in decompiler / lifter. warden.lift (lift_function, lift_module): a pure-Python stack-machine lifter that renders readable pseudo-C. CLI: warden lift <label> [--index N] [--out FILE]. It also backs warden export --format pseudo, so pseudocode export emits real pseudo-C instead of a mnemonic dump. Example: i32 parse_token(i32 p0, i32 p1) { return ((p0 + p1) * 7); }.
• Structured control-flow decompiler and round-trip symbol bridge. warden.lift now reconstructs structured control flow on top of the existing expression folding. Real if/else blocks are emitted (result-typed ifs assign each branch into a temp). While loops use the common block+loop idiom and render as while (1) { ... if (cond) break; ... } with no goto. A labeled goto is the fallback only for control flow that does not fit the innermost-loop/break pattern, so output stays correct. br_table becomes a switch. An unmodeled opcode degrades to a /* mnemonic */ comment, never a crash. Example outputs: abs_demo lifts to an if/else that returns; sum_to_n lifts to a while loop with break. The public API is unchanged (lift_function(module, func) -> str, lift_module(module) -> str) as is the CLI (warden lift <label> [--index N]). A new samples.control_flow() helper provides these two demo functions. warden.bridge adds a neutral-format round-trip for annotations. export_symbols(kb, version_id, fmt="csv") -> str (fmt: "csv" or "json") and import_symbols(kb, version_id, text, fmt="csv", *, provenance=None, confidence=None, lock=False) -> ImportResult (fields: matched, written, rejected_by_economy, skipped, unmatched, details). The import keys on the stable function identity first, then falls back to the function index, so a name recovered against one build lands on the same logical function in another build even when the index shifts. Imports pass through the provenance/confidence economy, so a file never clobbers higher-authority work. CLI: warden export --format csv|json (added alongside the existing headers, pseudo, kb-text, and ghidra formats) and a new warden import <label> <file> [--format csv|json] [--provenance human] [--lock] command.
• Agent crew. A propose, verify, write-back loop with a deterministic offline heuristic backend (no API key required), an optional OpenAI backend (gpt-5.3-codex by default, with codex and oai aliases), and an optional Anthropic backend (claude-opus-4-8 by default). LLM backends use structured JSON output.
• Call-graph agent strategy. warden.analysis.callgraph provides build_call_graph(module) (returns a CallGraph with .edges, .imports_called, .indirect_callers, .table_targets, and .callees(index)) and layered_schedule(module, graph=None) (bottom-up layers of function indices, with strongly-connected components condensed via iterative Tarjan). run_agent_pass gained two keyword arguments: strategy ("call-graph" by default, or "flat") and concurrency (int, default 8). CLI: warden agent <label> [--strategy call-graph|flat]. The call-graph strategy works in five steps: (1) build a static intra-module call graph (direct calls are exact; call_indirect / dynCall indirect calls are over-approximated to table targets of the matching type from module.elements); (2) condense SCCs and sort into bottom-up layers so every function’s defined callees are in earlier layers; (3) run the concurrency and struct analyzers first and route their findings into per-function notes (atomic sites, struct layouts); (4) process layers bottom-up, giving each function a FunctionFacts.callee_names list of its callees’ recovered names so the naming LLM sees callee meanings before producing a name for the caller; (5) functions within the same layer are independent and are proposed concurrently in-process via asyncio (blocking LLM backends run in worker threads capped by concurrency). Writes still go through the provenance/confidence economy, so concurrent branches sharing a callee cannot clobber each other. FunctionFacts gained two new fields: callee_names (list[str]) and notes (list[str]). The "flat" strategy preserves the original single-pass, leaves-first ordering.
• Specialized concurrency and struct analyzers. warden.analysis.concurrency (analyze_concurrency returns a ConcurrencyReport with .shared_memory, .atomic_sites, .pthread_markers, .facts) and warden.analysis.structs (analyze_structs returns StructLayout objects with .name, .fields, .source_function). Both populate the previously-empty thread_model and structs KB tables. CLI: warden analyze <label>.
• Verification. Determinism verification (runs today) plus a detected, optional wasm2c/w2c2 differential-equivalence plan. A zero-dependency interpreter (warden.interp: execute_function, differential_execute) makes behavioral-equivalence checking runnable without external tooling. CLI: warden exec <label> <index> [args...].
• Static HTML report generator. warden.report (render_report, write_report): a self-contained HTML file (inline CSS, no server required) with a coverage summary, a confidence heatmap colored by provenance and confidence, a thread/memory model section, and the diff changelog. CLI: warden report <label> [--out FILE].
• Exporters. C headers, readable pseudocode, a git-diffable KB text dump, and a Ghidra rename script.
• MCP server. Optional warden mcp tool surface mirroring the GhidraMCP pattern, including backend discovery, grounded function facts, server-side agent runs, and economy-gated symbol proposals.
• CLI and UX. warden init, ingest, versions, coverage, funcs, show, set-name, oracle, agent, agent-backends, diff, lift, exec, analyze, report, export, verify, mcp, and demo with rich terminal output.
• Test suite (247 tests), CI (lint, types, tests, wasm validation), a PyPI Trusted Publishing release workflow, a Mintlify documentation site, a Docker image, pre-commit hooks, and a full documentation set.