leptris 1.9.201.0-arm-linux → 1.9.201.2-arm-linux

data/vendor-src/libleptris/TODO.assemble/README.md

@@ -0,0 +1,16 @@
+# TODO.assemble — XInclude xml-mode, XPointer, xml:base, resolver
+
+Phases:
+- 01-xinclude-xml: parse="xml" splices the included document's
+  children (adopt/merge machinery, fresh-parse fallback, xi:fallback
+  honored, relative href resolution via xml:base chain, loop
+  detection, accept/accept-language passthrough stubs)
+- 02-xpointer: element() scheme (ID shorthand + child sequences),
+  xpointer() scheme (XPath selection, first-node-wins, points/
+  ranges explicitly out), xmlns() already shipped via ns bindings
+- 03-xml-base: base URI resolution on parse (xml:base attributes
+  merged per RFC 3986-ish relative resolution) + mutation updates
+- 04-resolver-policy: pluggable resolver hook (file/local default),
+  XML catalog support (parse + lookup), and the network policy —
+  remote fetch stays OFF unless explicitly enabled per-parse
+  (no_network default true; SSRF posture documented)

data/vendor-src/libleptris/TODO.bindings/01-mutation-construction.md

@@ -0,0 +1,18 @@
+# TODO.bindings/01-mutation-construction — Mutation/construction API
+
+From issue #510 (Tier 2/3 binding asks).
+
+leptris_document_new (alias of leptris_document_create), create element/text/comment/CDATA/PI, append/insert/remove/reparent, set text/attribute, subtree clone. Requires arena growth POST-PARSE (the pool is sized at parse time — growth is the core engineering risk). Unlocks the entire write path: lxml SubElement/append, Nokogiri builders, edit-then-serialize. Both bindings are read-only today only because this is missing. Rides into 02 (incremental building).
+
+Status: open. Design first (see the board README); one PR per item,
+full phases, like TODO.concurrency.
+DONE 2026-08-23: audit showed the full surface already shipped across
+earlier TODOs — create: leptris_document_create, element_create,
+text/comment/cdata/pi_node_create; wire: element_append_child,
+insert_before/after, remove_child/children, document_adopt_child,
+node_unlink; set: element_set_text, set_attribute (+int/uint/double/
+float/bool variants), remove_attribute(+all); clone: element_copy,
+document_copy, element_append/prepend_copy. The issue text predates
+this. What was missing was proof: DomBuilder round-trip spec (build
+from scratch -> serialize -> reparse -> verify, incl. all node types
++ batch children check) + deep-copy spec now in test_dom.cpp.

data/vendor-src/libleptris/TODO.bindings/02-incremental-iterparse.md

@@ -0,0 +1,15 @@
+# TODO.bindings/02-incremental-iterparse — Incremental tree building / true iterparse
+
+From issue #510 (Tier 2/3 binding asks).
+
+Builder + push-parse + subtree prune/free on the arena from 01. Bounded-memory processing of huge files — a top lxml selling point. Subtree free must return nodes to the pool (or a dead-list) without invalidating siblings.
+
+Status: open. Design first (see the board README); one PR per item,
+full phases, like TODO.concurrency.
+DONE 2026-08-23: leptris_iterparse_new/_next/_free over the pull API.
+Each top-level child of the root materializes in its OWN document
+(pool); the next _next() releases the previous subtree document —
+memory bounded by the largest subtree, not the document. Text runs
+accumulate across pull text events before node creation. v1
+limitation documented in the header: QNames as written, prefixes not
+re-resolved. Specs: test/sax/test_pull.cpp (Iterparse.*).

data/vendor-src/libleptris/TODO.bindings/03-compiled-xpath.md

@@ -0,0 +1,15 @@
+# TODO.bindings/03-compiled-xpath — Compiled XPath expressions
+
+From issue #510 (Tier 2/3 binding asks).
+
+leptris_xpath_compile/eval/free with a stated thread contract (the AST/bc cache is process-global + mutex/pin-guarded today — a compiled handle pins its entry for its lifetime). Hot-loop win on top of the batch accessors (#509): today every leptris_xpath_eval re-checks the cache by string hash.
+
+Status: open. Design first (see the board README); one PR per item,
+full phases, like TODO.concurrency.
+DONE 2026-08-23: leptris_xpath_compile / _compiled_eval /
+_compiled_free wrapping the pinned AST cache — one parse + one pin
+for the handle's lifetime, bytecode cached on first eval; skips the
+per-call hash + cache probe of leptris_xpath_eval. Thread contract:
+immutable handle, concurrent eval safe (pinned entry), free after
+last eval; failures snapshot into the document error slot. Specs:
+test/xpath/test_compiled.cpp incl. a 4-thread one-handle stress.

data/vendor-src/libleptris/TODO.bindings/04-pull-stax-api.md

@@ -0,0 +1,14 @@
+# TODO.bindings/04-pull-stax-api — Pull (StAX-style) API over the SAX core
+
+From issue #510 (Tier 2/3 binding asks).
+
+next_event instead of callbacks. C→host callbacks cost ~µs each through FFI and eat the streaming advantage; a host-driven pull loop keeps it. Reuses sax/parser.c states; add an event struct + cursor handle.
+
+Status: open. Design first (see the board README); one PR per item,
+full phases, like TODO.concurrency.
+DONE 2026-08-23: leptris_pull_new/_next/_free + attr accessors over
+the streaming SAX machine. Events queue from 256-byte input slices
+fed on demand; strings owned until the next _next. ERROR event
+carries the message; END_DOCUMENT terminates success walks. Specs:
+test/sax/test_pull.cpp (order, attrs, comment/CDATA/PI, malformed,
+invalid input).

data/vendor-src/libleptris/TODO.bindings/05-per-parse-options.md

@@ -0,0 +1,17 @@
+# TODO.bindings/05-per-parse-options — Per-parse options struct
+
+From issue #510 (Tier 2/3 binding asks).
+
+Replace global leptris_set_strict_mode (global mutable state is thread-hostile): entity resolution mode, depth/size/entity-expansion caps, whitespace policy, XInclude fetch policy. The XInclude no_network flag matters — external fetching is an SSRF/filesystem hazard that deserves an explicit opt-in.
+
+Status: open. Design first (see the board README); one PR per item,
+full phases, like TODO.concurrency.
+DONE 2026-08-23: LeptrisParseOptions {flags, strict_mode, max_depth}
++ leptris_parse_string_ex — save/restore of the thread-local
+defaults around one call (documented not-reentrant). The XInclude
+no_network flag was dropped from v1: the resolver never fetches over
+the network (file/local only), so the hazard it guards does not
+exist today — revisit if/when remote XInclude lands. Global setters
+remain for compatibility. Specs: ParseOptions.* in test_parser.cpp
+(depth cap + restore, defaults == plain parse, flags passthrough).
+LeptrisParseFlags moved to types.h (single canonical types source).

data/vendor-src/libleptris/TODO.bindings/06-serialize-encoding-guarantees.md

@@ -0,0 +1,14 @@
+# TODO.bindings/06-serialize-encoding-guarantees — Serialize-with-encoding round-trip guarantees
+
+From issue #510 (Tier 2/3 binding asks).
+
+Document what output encodings are promised across the iconv-optional build matrix (LEPTRIS_ENABLE_ICONV on/off × input encodings), and what serialize does when the target encoding is unavailable. Bindings need the contract, not per-platform probing.
+
+Status: open. Design first (see the board README); one PR per item,
+full phases, like TODO.concurrency.
+DONE 2026-08-23: contract fixed and documented. Output is ALWAYS
+UTF-8 (iconv is input-side only); the XML declaration never lies —
+docs parsed as ISO-8859-1 (or explicit non-UTF-8 serialize requests)
+now declare UTF-8 instead of the pre-fix mismatch (declaration said
+ISO-8859-1 while the body was UTF-8). serialize(parse(serialize(x)))
+byte-stable, spec'd. README "Serialization encoding guarantees".

data/vendor-src/libleptris/TODO.bindings/README.md

@@ -0,0 +1,36 @@
+# TODO.bindings — binding-driven engine asks (issue #510)
+
+Tracked asks from the lxml-mirror Python binding (leptris-py) and the
+Nokogiri-mirror Ruby binding, ranked by binding value ÷ effort.
+
+## Shipped (Tier 1 / Tier 3, v1.3.0)
+
+- `leptris_version()` / `leptris_version_components()` — runtime
+  version detection at load time.
+- Parse-error positions: `leptris_last_error_position(line, column)`
+  next to `leptris_last_error()`.
+- `leptris_node_line(node)` — 1-based source line, on demand.
+- Batch children fetch: `leptris_element_children` (issue #509);
+  result batch: `leptris_xpath_result_get_nodes(_ex)`.
+- Documented threading contract: README "Threading model"
+  (one-document-per-thread; TLS error channel; `leptris_thread_cleanup`).
+- Header/binary drift gate: export-surface CI check (issue #508).
+
+## Executed (v1.4.0) — see DONE notes in each file
+
+- 01 — mutation/construction API: audit found the full surface
+  already shipped; added the builder round-trip + deep-copy specs
+- 02 — incremental tree building / true iterparse:
+  leptris_iterparse_* (per-subtree pools, bounded memory)
+- 03 — compiled XPath expressions: leptris_xpath_compile /
+  _compiled_eval / _compiled_free (pinned cache entry)
+- 04 — pull (StAX-style) API: leptris_pull_* over the streaming
+  SAX core (no C->host callbacks)
+- 05 — per-parse options: LeptrisParseOptions +
+  leptris_parse_string_ex (scoped strict/depth/flags)
+- 06 — serialize-encoding guarantees: UTF-8 output always,
+  truthful declarations, byte-stable double-serialize
+
+## Non-goals (keep leptris small)
+
+XSLT, XSD/RelaxNG/DTD validation, HTML parsing, XPath 2.0+.

data/vendor-src/libleptris/TODO.concurrency/01-thread-safe-errors.md

@@ -0,0 +1,19 @@
+# 01 — Thread-safe error reporting + per-document error slots
+
+User report (leptris-ruby): `leptris_last_error` returns a
+process-global string — races under concurrent parses from threaded
+apps.
+
+- error.c: `error_message` / `last_error` become LEPTRIS_THREAD_LOCAL
+  (fixes the race under the one-document-per-thread contract).
+- `struct leptris_document` gains a `last_error_message[256]` slot;
+  every public parse entry point snapshots the thread-local message
+  into it on failure.
+- New public API: `leptris_document_last_error(doc)` — the message
+  from this document's most recent failing parse (NULL when none or
+  on NULL doc). Doc-owned; lives until leptris_document_free.
+- `leptris_last_error` stays (legacy), now documented as the
+  best-effort THREAD-LOCAL last message.
+
+DONE 2026-08-23: all four landed; specs cover per-doc slots under
+concurrent parses and the documented legacy semantics.

data/vendor-src/libleptris/TODO.concurrency/02-export-surface-gate.md

@@ -0,0 +1,24 @@
+# 02 — Exported-but-undeclared symbols: close the gap, gate it
+
+User report: the internal `leptris_parse` family (src/leptris/
+leptris_parse.h) exports from the dylib yet appears in no public
+header — invisible to header-derived bindings. Root cause: non-Windows
+builds use default visibility, exporting EVERY internal symbol; only
+Windows is narrow (dllexport-only).
+
+- Build: `C_VISIBILITY_PRESET hidden` on leptris_objects — internal
+  symbols stay resolvable inside the final binary (tests, CLI embed
+  the objects) but no longer export from the shared library.
+  LEPTRIS_API (visibility default) overrides for the public surface.
+- Decision on `leptris_parse` family: UN-EXPORT, do not promote — it
+  returns `struct leptris_document*` and takes no status out-param;
+  promoting would leak internal shapes. Bindings use
+  `leptris_parse_string*` / `leptris_parse_file*`.
+- New gate: scripts/check_export_surface.py — nm -gU on the shared
+  library diffed against LEPTRIS_API declarations across
+  src/include/. Wired as a ctest (skips when no shared lib) and a
+  CI job with a shared build. Same class as the #468 retro-declared
+  leptris_serialize_document.
+
+DONE 2026-08-23: visibility preset + un-export + gate script + CI job;
+gate diff is empty (every exported symbol is declared public).

data/vendor-src/libleptris/TODO.concurrency/03-mixed-nodeset-batch.md

@@ -0,0 +1,14 @@
+# 03 — Mixed-nodeset batch accessor
+
+User report (#477 residual): `leptris_xpath_result_get_nodes` copies
+2-of-4 on a mixed nodeset — the elements-only contract is real but
+undocumented, forcing bindings into per-index fallbacks.
+
+- Document `leptris_xpath_result_get_nodes` as ELEMENTS-ONLY (header
+  doc block).
+- New API: `leptris_xpath_result_get_nodes_ex(result, out_nodes,
+  out_kinds, max_count)` — copies EVERY entry plus its
+  LeptrisXPathNodeKind; returns the copied count.
+
+DONE 2026-08-23: both landed; spec asserts 4-of-4 with kinds on the
+mixed nodeset that previously truncated.

data/vendor-src/libleptris/TODO.concurrency/04-canonical-status-message.md

@@ -0,0 +1,13 @@
+# 04 — One canonical status-message function
+
+User report: `leptris_status_string` and `leptris_error_message`
+have overlapping contracts; bindings guess which is richer. Since the
+phantom implementation landed, error_message delegates to
+status_string — identical output.
+
+- Canonical: `leptris_status_string`.
+- `leptris_error_message` documented as a deprecated-but-stable
+  alias in BOTH doc blocks (kept: removing it breaks the Ruby/Python
+  mirrors already shipping).
+
+DONE 2026-08-23: doc blocks updated on both.

data/vendor-src/libleptris/TODO.concurrency/05-windows-dll-name.md

@@ -0,0 +1,10 @@
+# 05 — Windows DLL naming
+
+User report: the DLL is `leptris.dll` while every other platform
+produces `libleptris.*`; vendoring tooling special-cases it.
+
+- CMake: on WIN32, leptris_shared gets OUTPUT_NAME `libleptris`
+  (elsewhere OUTPUT_NAME stays `leptris` → lib prefix does the job).
+
+DONE 2026-08-23: conditional OUTPUT_NAME; Windows now produces
+libleptris.dll. check_dll_exports.py updated to accept the new name.

data/vendor-src/libleptris/TODO.concurrency/06-exslt-pack.md

@@ -0,0 +1,26 @@
+# 06 — First-party EXSLT-style extension pack
+
+User report: `leptris_xpath_register_function` exists precisely for
+this — ship one C implementation of the commonly-missed XPath 2.0
+conveniences instead of N binding reimplementations.
+
+Public API: `leptris_exslt_enable(doc)` — registers the pack for that
+document (scoped, freed with the doc; prefixed names can never clash
+with the XPath 1.0 core). Native handlers (full result types), not
+the string-only custom-fn bridge:
+
+- str: replace, tokenize, split, concat, padding
+- set: distinct, intersection, difference, leading, trailing
+- math: max, min, abs, sqrt, power
+
+Scope notes: str:replace takes string args (EXSLT's parallel-nodeset
+form is out of scope v1); tokenize/split build nodesets of synthetic
+text nodes (XPathNodeSet grows an owns_synthetic_text flag to free
+them, mirroring owns_attributes).
+
+DONE 2026-08-23: 15 handlers + enable API + specs for every function
+incl. kind checks on tokenize output; registry lookup by prefixed
+name (parser stores "str:replace" for QName tokens). Semantic split
+ honored: str:tokenize matches a CHARACTER SET (any delimiter char),
+str:split matches the whole PATTERN as a substring
+("a::b" split on "::" = a|b, not a||b).

data/vendor-src/libleptris/TODO.concurrency/07-ns-set-pairs.md

@@ -0,0 +1,11 @@
+# 07 — Array-based ns-set constructor
+
+User report: ns_set_new + N×_add is N+1 FFI calls per query. A flat
+array constructor makes it one — same wire shape as the c14n
+inclusive-namespaces argument, so bindings share their adapter.
+
+- New API: `leptris_xpath_ns_set_new_from_pairs(const char* const*
+  flat, size_t pair_count)` — flat = [p1, u1, p2, u2, ...].
+  NULL/empty-prefix entries rejected (LEPTRIS_ERROR_NULL_ARG).
+
+DONE 2026-08-23: landed; spec covers 3-pair construction + rejection.

data/vendor-src/libleptris/TODO.concurrency/08-threading-contract.md

@@ -0,0 +1,45 @@
+# 08 — The threading contract, stated and enforced
+
+The concurrency audit (2026-08-23): documents are independent (pool,
+element index, rank cache, custom-fn registrations are per-document);
+the compact overflow table and current-document pointer are already
+thread-local. Remaining process-global hazards:
+
+- error.c channel → fixed by 01 (thread-local).
+- XPath AST/bytecode cache (xpath_ast_cache.c g_cache) — stays
+  PROCESS-GLOBAL under a mutex (a TLS cache would leak ASTs on
+  thread exit; no portable C99 TLS destructors). Three races fixed:
+  (a) twin insert freed the caller's AST while it was still in use —
+  insert now returns the canonical (pinned) AST; (b) LRU eviction
+  freed AST/bytecode underneath a running evaluate — entries are now
+  pin-counted and evicted-but-pinned entries go to a graveyard freed
+  on last release; (c) compile reads the shared AST, verified
+  read-only.
+- g_standard_registry lazy init — double-init race on first
+  concurrent eval; now built under a dedicated mutex with every read
+  through the same lock (a load-time constructor was tried first but
+  reverted: it allocated the registry in every linked process, even
+  XPath-free ones, and valgrind flagged it in LTO binaries where the
+  global became unreachable).
+- SIMD dispatch pointers (simd_text.c) and the parser's cached
+  getenv debug flags — lazy first-call writes raced; dispatch now
+  constructor-initialized, the getenv flags are thread-local.
+
+Deliverable: README "Threading" section stating the contract — one
+document per thread needs no locking; documents may migrate between
+threads only between calls; nothing else is shared. Plus a
+concurrency stress spec (test/concurrency): 4 threads ×
+parse/xpath/serialize/free loops, ASAN-clean.
+
+DONE 2026-08-23: README "Threading model" section + stress spec
+(test/concurrency: 4 threads x parse/eval/serialize/free, TLS error
+isolation, per-doc error slots). Found and fixed one real corruption
+(twin-insert use-after-free, reproduced: first eval per thread
+returned "Unsupported AST node type: <garbage>") and two TSAN races
+(SIMD dispatch init, getenv flag caching). ThreadSanitizer-clean on
+the full suite; the lone TSAN failure (CliFormat.CompactByDefault)
+is the macOS TSAN-runtime-after-fork artifact, not a library race.
+Follow-ups caught by Linux CI: per-thread free lists (XPath results/
+nodesets, root-map entries) retained memory after thread exit — new
+public leptris_thread_cleanup() drains them from each worker; the
+registry went mutex-lazy (see above).

data/vendor-src/libleptris/TODO.engine/01-file-sources.md

@@ -0,0 +1,11 @@
+# TODO.engine/01-file-sources — File-backed pull and iterparse sources
+
+leptris_pull_new_file(path) / leptris_iterparse_new_file(path): the puller reads the input in PULL_SLICE chunks straight from a FILE* — huge documents stream off disk with memory bounded by the slice + one subtree (iterparse), no whole-document buffer. The memory constructors stay.
+
+DONE 2026-08-24: leptris_pull_new_file(path) + leptris_iterparse_new_file(path).
+The puller gained a FILE* source — slices are fread into a bounded
+256-byte buffer, so both constructors stream off disk with no
+whole-document buffer (iterparse stays bounded by the largest
+subtree). Memory constructors unchanged. Specs: Pull.FileSourceMatchesMemorySource
++ Iterparse.FileSourceYieldsSubtrees (every event kind; missing-file
+NULL). README notes the file variants.

data/vendor-src/libleptris/TODO.engine/02-compiled-contexts.md

@@ -0,0 +1,10 @@
+# TODO.engine/02-compiled-contexts — Compiled XPath with ns and variable contexts
+
+leptris_xpath_compiled_eval_ns / _eval_vars: the compiled handle on the context-carrying paths. The ns/vars routes run the direct evaluator (the VM fast paths skip prefixed name tests), so these variants evaluate the pinned AST with ns_set / variable_set installed on the context — same semantics as leptris_xpath_eval_ns / eval_with_vars_context, minus the re-parse.
+
+DONE 2026-08-24: leptris_xpath_compiled_eval_ns / _eval_vars via a shared
+compiled_eval_context helper — the pinned AST evaluated directly
+(VM fast paths skip prefixed name tests, matching the plain-path
+design) with ns_set / variable_set installed on the context; doc-slot
+error snapshotting preserved. Specs: CompiledXPath.EvalWithNamespaceBindings,
+EvalWithVariables. README documents the variants.

data/vendor-src/libleptris/TODO.engine/03-rust-publish.md

@@ -0,0 +1,9 @@
+# TODO.engine/03-rust-publish — Rust crate publishing (crates.io)
+
+bindings/rust builds and tests in CI but has no publish path. Add a release workflow (workflow_dispatch + workflow_call, mirroring the leptris-py pattern) that publishes to crates.io with the CARGO_REGISTRY_TOKEN secret. crates.io has no OIDC trusted publishing — the secret is the mechanism; document the one-time setup.
+
+DONE 2026-08-24: .github/workflows/rust-release.yml (workflow_dispatch +
+workflow_call): builds the C core, cargo test, then cargo publish —
+packs with a warning when CARGO_REGISTRY_TOKEN is absent (one-time
+setup: crates.io token, publish-new scope, repo secret). Crate version
+bumped to engine lockstep 1.4.0 (was 1.1.1).

data/vendor-src/libleptris/TODO.engine/04-ruby-release-input.md

@@ -0,0 +1,8 @@
+# TODO.engine/04-ruby-release-input — leptris-ruby release bump computes major/minor/patch
+
+Their release.yml accepts an explicit x.y.z only — the advertised major/minor/patch values are not computed from the current version. Read LIBLEPTRIS_VERSION (Rakefile) + gem version and compute the bump in the workflow, mirroring the leptris C-repo bump-version.sh contract.
+
+DONE 2026-08-24: verified already fixed upstream — leptris-ruby's
+release.yml bump job accepts x.y.z OR major/minor/patch (computed
+from Leptris::VERSION, with a semver guard). No change needed;
+TODO.md note updated.

data/vendor-src/libleptris/TODO.engine/05-todo-md-refresh.md

@@ -0,0 +1,7 @@
+# TODO.engine/05-todo-md-refresh — TODO.md refresh
+
+The leptris-py 'pending trusted-publisher setup' note is stale — wheels have shipped repeatedly (1.3.x, 1.4.x). Update TODO.md to reflect reality and point the Rust items at the new workflow.
+
+DONE 2026-08-24: TODO.md updated — leptris-py trusted-publisher note
+closed (wheels shipping since 1.3.x), Rust item points at the new
+workflow + remaining one-time secret, ruby bump note closed.

data/vendor-src/libleptris/TODO.engine/README.md

@@ -0,0 +1,16 @@
+# TODO.engine — post-v1.4.0 improvement board
+
+Gaps surfaced while shipping TODO.concurrency (v1.3.0) and
+TODO.bindings (v1.4.0), plus the stale items in TODO.md.
+
+- 01 — file-backed pull/iterparse sources (true disk streaming)
+- 02 — compiled XPath with ns/variable contexts
+- 03 — Rust crate publishing workflow (crates.io)
+- 04 — leptris-ruby release bump: compute major/minor/patch
+- 05 — TODO.md refresh (stale leptris-py trusted-publisher note)
+
+All five shipped 2026-08-24 (DONE notes in each file).
+
+Deliberately NOT boarded (stay in TODO.md): the document-node model
+(XPath conformance edge — a DOM-model project, not an increment),
+XInclude parse="xml"/xpointer (TODO 92), DTD phase 8 (TODO 91).

data/vendor-src/libleptris/TODO.hygiene/README.md

@@ -0,0 +1,15 @@
+# TODO.hygiene — conformance completeness
+
+Phases:
+- 01-document-node: XPath-level synthetic document node — //node(),
+  /, count(/), and the generic descendant expansion select/see the
+  document as a node without a DOM overhaul (nodesets carry a
+  sentinel; accessors handle it)
+- 02-dtd-phase8: ENTITY/ENTITIES attribute resolution, parameter
+  entities in the internal subset, choice-model backtracking in
+  content validation (TODO 91 closeout)
+- 03-ns-fixup-mutation: namespace fixup on adopt/insert across
+  parents (undeclared prefixes repaired by declaration synthesis,
+  per DOM Level 3 semantics)
+- 04-spec-sweep: every new behavior falsifiable — violation cases
+  required per spec group (project rule)

data/vendor-src/libleptris/TODO.max-perf/00-INDEX.md

@@ -0,0 +1,22 @@
+# TODO.max-perf — remaining-work ledger, lowest-hanging first
+
+Rank = impact ÷ effort. Work top-down; one lever per PR with a
+RED bench first where perf claims are made.
+
+| # | Lever | Type | Effort | Status |
+|---|-------|------|--------|--------|
+| 3 | iterparse fixed-cost (6.9µs) + error channel (#563) | perf | bench landed #1195; remaining = zero-copy one-shot emission (design with lever 2) | half done |
+| 1 | SIMD/memchr HTML tokenizer scans | perf | memchr ×2 (#1176/#1190) + LUT (#1199, +30%/+51%) | DONE v1.9.199 |
+| 7 | leptris validate CLI + bindings | feature | engine half merged #1198 (--dtd + DTD engine fixes); bindings remain | half done |
+| 8 | native XML diff consumer surface | feature | small | next |
+| 4 | compact-node micro-opts (non-layout) | perf | audited: TODOs 172/195/Lane18-P3 already in; residual = node-creation cost (node-layout fork, owner call) | parked on fork |
+| 2 | compiled-dispatch tokenizer | perf | medium; design together with lever 3's zero-copy emission | queued |
+| 5 | XSLT 2x per-bench profile rounds (#682) | perf | medium, ongoing | in progress |
+| 6 | language tails (QT3/dates/fn-items/lane 15) | correctness | medium-large | queued |
+| 9 | prebuilt-gem 12-platform matrix | release | small, BLOCKED on owner go | blocked |
+| 10 | moxml Plan follow-ups (upstream) | upstream | small | asks pending |
+
+Portability wave 2026-09-18 (v1.9.199): ILP32 layout port #1193 —
+per-wordsize pins, SSE2 gate, xsl:number 64-bit chain, ILP32 +
+native-arm64 + Intel-macOS CI legs, concurrency groups everywhere.
+Next portability gate: s390x big-endian (#1194, PR #1202).

data/vendor-src/libleptris/TODO.max-perf/1-simd-html-scanner.md

@@ -0,0 +1,15 @@
+# Lever 1: SIMD byte classification for the HTML tokenizer hot loops
+
+Status: TODO. Research (2022-2026): simdjson/Lemire (arXiv:2503.01662)
+shows GB/s lexing from SIMD byte classification; our
+`common/simd_text.c` already has dispatch hooks (`g_contains_fn`,
+`g_find3_fn`, `g_count_fn`, `g_scan_fn`) to bolt implementations in
+without touching call sites.
+
+Levers:
+- Vectorized tag-open/quote/ws/name-class scans in the html_parse.c
+  hot loops (`h_pooled_lower`, attr scanner, name/ws scans).
+- Replace per-byte `while (q < end && !h_is_ws(*q))` scans with
+  the dispatched `g_find*` primitives.
+Gate: `benchmarks/dom_benchmark` vs pugixml attr-parse row (#682);
+fresh build dirs, identical CMAKE_BUILD_TYPE (feedback_benchmark_discipline).

data/vendor-src/libleptris/TODO.max-perf/10-moxml-plan-followups.md

@@ -0,0 +1,9 @@
+# Lever 10: moxml Plan follow-ups (upstream asks)
+
+Status: Plan is merged upstream (lutaml/moxml#251, 0.5.60).
+Asks for the next round:
+- Promote plan_rows to a documented adapter contract + conformance
+  spec (base adapter currently nil-fallbacks to the NodeSet walk).
+- Port plan_rows to the other adapters.
+- Add a Plan row to the moxml bench matrix (lock the
+  raw-nokogiri-parity claim in CI).

data/vendor-src/libleptris/TODO.max-perf/2-compiled-dispatch-tokenizer.md

@@ -0,0 +1,8 @@
+# Lever 2: compile-don't-interpret the tokenizer dispatch
+
+Status: TODO. Precedents: XGrammar-2 JIT + tag-dispatched switching
+(arXiv:2601.04426), tree-sitter grammar-to-C, CPython's generated
+PEG. Our html_parse.c start/end-tag ladder is a long strcmp chain
+per token; replace with a precomputed first-char dispatch table
+(built once per process) + name-hash switch, like the XPath VM's
+BYTE_DISPATCH lane. Same pattern applies to the end-tag scope walk.

data/vendor-src/libleptris/TODO.max-perf/3-iterparse-fixed-cost.md

@@ -0,0 +1,5 @@
+# Lever 3: #563 iterparse 6.9us fixed cost + error channel
+
+Status: TODO (task #3). Per-call cost dominated by re-setup; cache
+the scan environment per document (the exec-scoped eval-environment
+pattern from task #4 applies), batch events (#585/#589 machinery).

data/vendor-src/libleptris/TODO.max-perf/4-compact-node-micro-opts.md

@@ -0,0 +1,17 @@
+# Lever 4: Lane 18 compact-node micro-optimizations (non-layout)
+
+Status: TODO. Lane 18 gate: element append ≤ pugixml, attr-parse
+≤ pugixml. PR #1017 closed attr-parse to 1.96x; the mutation rows
+need the node-layout redesign (parked — NOT this lever). This
+lever is the NON-layout remainder.
+
+Low-hanging:
+- append_child: sibling-tail cache on the parent (O(1) append
+  instead of chain walk) if not already present.
+- attr lookup: interned-name pointer compare before strcmp in
+  leptris_element_attr (names are pooled; identical strings are
+  common in real docs).
+- pool alloc batching for text nodes (arena bump is already O(1);
+  batch the header init).
+
+RED bench first: benchmarks/dom_benchmark append + attr rows.

data/vendor-src/libleptris/TODO.max-perf/5-xslt-bench-profiling-rounds.md

@@ -0,0 +1,13 @@
+# Lever 5: #682 XSLT 2x — per-bench profile rounds
+
+Status: TODO. Gate: every XSLT bench ≥2x libxslt and Saxon.
+Lane is in progress; the remaining gap benches need individual
+profiling rounds (perf record on the specific bench, top-5
+self-time, one lever each).
+
+Rounds:
+1. List benches below 2x from the last scorecard run.
+2. For each: perf record, identify the top self-time function,
+   land ONE lever, re-measure.
+Candidates seen in past rounds: pattern match linear scans,
+output escaping per-byte, variable lookup re-walks.

data/vendor-src/libleptris/TODO.max-perf/6-language-tails.md

@@ -0,0 +1,9 @@
+# Lever 6: Stage-2 language tails (QT3 subset, dates, fn-items, lane 15)
+
+Status: TODO. Correctness work, not perf. Lanes 11/12 QT3 subset
+adoption; date/time function family; fn-items family; lane 15
+miscellaneous XQuery tails.
+
+Order: QT3 subset (lanes 11/12, corpus-driven, largest) →
+dates → fn-items → lane 15 stragglers. Each = adopt the corpus
+slice, RED cases first, one PR per family.

data/vendor-src/libleptris/TODO.max-perf/7-validate-bindings.md

@@ -0,0 +1,6 @@
+# Lever 7: Lane 16.6 — leptris validate CLI + bindings
+
+Status: TODO. `leptris validate` subcommand (XML well-formedness
++ DTD + RNG paths through the public API) and the binding-side
+validate wrappers. Small, closed scope; CLI command registration
+follows cli/CLI_ARCHITECTURE.md (new file in cli/commands/).

data/vendor-src/libleptris/TODO.max-perf/8-native-xml-diff-consumer.md

@@ -0,0 +1,5 @@
+# Lever 8: Lane 17 — native XML diff consumer
+
+Status: TODO. The diff engine exists (lane 17 core); remaining is
+the consumer surface: CLI output formatting (patch/summary modes)
+and the binding wrapper. Closed scope.

data/vendor-src/libleptris/TODO.max-perf/9-prebuilt-gem-matrix.md

@@ -0,0 +1,14 @@
+# Lever 9: prebuilt-gem platform matrix — parsanol-parity 12
+
+Status: BLOCKED on the owner's go for the platform SET (releases
+are the owner's call). Target (identical across leptris/yeptris/
+parsanol): ruby, x86_64-linux, x86_64-linux-musl, aarch64-linux,
+aarch64-linux-musl, arm-linux, arm-linux-musl, x86_64-darwin,
+arm64-darwin, x64-mingw32, x64-mingw-ucrt, aarch64-mingw-ucrt.
+
+Rationale (consumer-side): many Ruby users cannot compile; a
+missing platform gem is a hard install failure. yeptris is
+missing x86_64-linux-musl, aarch64-linux-musl, x64-mingw32 today.
+
+Effort once approved: release.yml matrix edits + one release per
+gem. Half a day.

data/vendor-src/libleptris/TODO.md

@@ -0,0 +1,33 @@
+# Open work
+
+Everything from the TODO.fix (110 executed plans) and TODO.remaining
+(10 closed items) boards has shipped or was closed as measured-dead —
+those files are removed from the tree and preserved in git history
+(release v1.1.2-era `main` has the full set). This file tracks what
+is genuinely still open, with pointers.
+
+## Library
+
+- **Document node** — the engine has no document node. `//node()`
+  selects the root element via the folded absolute walk, but the
+  generic expansion (`/descendant-or-self::node()/child::node()`
+  written explicitly) cannot; a real document-node model is the
+  eventual fix. Also the boundary for `count()`-style expressions
+  written out longhand.
+
+## Bindings
+
+- **Rust: serde interop** — deferred until a consumer asks for it
+  (TODO.remaining/05 closure note, git history).
+- **Rust: crates.io publishing** — publish workflow shipped
+  (.github/workflows/rust-release.yml, TODO.engine/03); the crate
+  version rides engine lockstep. One-time secret remains: set
+  CARGO_REGISTRY_TOKEN (crates.io token, publish-new scope) and
+  dispatch — without it the workflow packs and warns.
+- **leptris (py)** — own repo, own pin, wheels shipping since 1.3.x
+  (trusted publishing is set up; releases dispatch from leptris-py).
+
+## Adjacent repos
+
+- **leptris-ruby** — release bump now computes major/minor/patch
+  from the current version (fixed in their release.yml).