fix(deps): update module github.com/odvcencio/gotreesitter to v0.53.0

[renovate]

This PR contains the following updates:

Package	Change	Age	Confidence
github.com/odvcencio/gotreesitter	v0.52.0 → v0.53.0

---

Release Notes

odvcencio/gotreesitter (github.com/odvcencio/gotreesitter)

v0.53.0: gotreesitter v0.53.0

Compare Source

Release overview

• This release fixes public API contract faults and C-parity gaps that a
  repository audit found. It also includes the maintenance work merged after
  v0.52.0.
• Each parse applies its timeout once. Tree handles are safe to release in
  the C order. Incremental reuse is correct for edit sequences and changed
  included ranges.
• The default memory budget grows with the input size, so valid large inputs
  no longer stop early.
• Reserved words, query predicates, and generated grammar tables match C in
  more cases.
• Eligible fresh parses use the compact parser by default. Unsupported cases
  retain the legacy fallback. This release does not complete compact parser
  graduation.

Performance evidence

Paired randomized benchmarks compare v0.52.0 with the v0.53.0 candidate code at
48503fef. The run used 20 shuffle seeds, alternating order, -benchtime=750ms,
GOMAXPROCS=1, and the gts_parsercorephase0 tag. It ran in the harness
container with 4 GiB of memory and one pinned CPU (Intel Core Ultra 9 285).

Benchmark	v0.52.0	v0.53.0	Change
BenchmarkGoParseFullDFA	13.099 ms, 258.6 KiB, 42 allocs	9.637 ms, 226.5 KiB, 37 allocs	-26.43% time
BenchmarkGoParseIncrementalSingleByteEditDFA	2,294.1 us, 184.2 KiB, 95 allocs	127.5 us, 3.5 KiB, 5 allocs	-94.44% time
BenchmarkGoParseIncrementalNoEditDFA	2.612 ns, 0 allocs	4.043 ns, 0 allocs	+54.80% time

All three time changes have p=0.000 with n=20.

• The single-byte edit gain comes from the authenticated token-invariant
  reuse that v0.52.0 disabled and that returned after it.
• The no-edit regression is 1.4 ns. The unchanged-tree fast path now adds a
  tree handle and compares included ranges. It stays in single-digit
  nanoseconds with no allocations. Recover it in a later release.
• Each returned tree now allocates one new Tree value, because released
  trees no longer return to a pool.

A one-shot large-file run parses the canonical grammargen/lr.go fixture with
BenchmarkParityGoCanonicalFull under /usr/bin/time -v in the same container.

Measure	v0.52.0	v0.53.0
Maximum resident set size	151,444 KiB	137,540 KiB
Bytes allocated for each parse	4,304,096 B	134,560 B
Allocations for each parse	64,132	3,080

The run used GOMAXPROCS=1 and -benchtime=1x. The raw outputs stay outside the repository.

Parse timeout

• Apply SetTimeoutMicros once for each parse. The compact route and the production fallback now share one deadline.
• A parse that the compact route stopped on a timeout ran for about twice the configured timeout before this change.

Tree handles

• Make a second Release on a released tree do nothing. A stale call no longer frees the tree of a later parse.
• Stop pooling Tree values. Each returned tree now costs one new 3,296-byte value.
• Add a handle when an unchanged incremental parse returns its old tree. Releasing the old tree no longer invalidates the result.
• Document that ParseIncremental updates parent links in the old tree, and that Node.Edit after Tree.Edit moves spans twice.

Incremental edit sequences

• Clear dirty nodes on byte-identical source only when the recorded edits restore every node span. A delete-then-reinsert sequence no longer reuses a collapsed leaf.
• Parse again from the start when the parser included ranges differ from the old tree ranges.
• Compare random multi-edit sequences with fresh parses for JSON and Go.

Memory budget

• Scale the default per-parse memory budget with the input. It is the larger of 512 MiB and 512 bytes for each input byte. Valid 7 MB JSON no longer stops with ParseStopMemoryBudget.
• Keep the process-heap ceiling at the larger of 2 GiB and twice the budget.
• Add Parser.SetMemoryBudgetBytes, Parser.MemoryBudgetBytes, and WithParserPoolMemoryBudgetBytes for a fixed budget.

Reserved words

• Promote a reserved word to its keyword token, as C ts_parser__lex does. JavaScript var if = 1; now reports an error.
• Attach reserved-word tables for JavaScript, OCaml, PHP, Pkl, Python, and templ from generated sidecars. Blob hashes do not change.
• Add the ts2go -reservedwords-only sidecar mode.

Query predicates

• Apply text predicates to every node of a quantified capture. The any- predicates need one matching node.
• Reject a predicate that names a capture the query has not bound, as C does.
• Keep matching unchanged after DisableCapture. The returned match only omits the disabled capture.

Generated grammar tables

• Encode shift targets with 32 bits in action-group keys. Grammars with more than 65,535 states no longer merge unrelated shift actions.
• Expand case-insensitive character-class ranges without extra characters.
• Reject a production with more than 255 right-hand-side symbols.

Continuous integration gates

• Fail continuous integration when a workflow -run pattern names a test that does not exist. Correct seven stale names.
• Run the exhaustive 206-language parity sweep every night.
• Require the build check for merges to main. Administrators can still bypass it.

Stack hashing

• Pack node flags with masks and shifts without changing hash values.

Generated CSS token precedence

• Preserve the authored precedence of named immediate tokens.
• Keep longer preferred tokens reachable after an immediate token accepts a prefix.
• Parse escaped CSS unit suffixes without introducing error nodes.

Generated HCL splat expressions

• Preserve right associativity when a proven repeat continuation has equal precedence.
• Keep attribute and index chains inside HCL splats in nested blocks.
• Compare generated splat trees with the locked C grammar.

Generated supertype aliases

• Use public alias symbols in generated supertype maps.
• Remove duplicate subtype entries after alias resolution.
• Test named aliases, anonymous name collisions, query captures, and the locked C map for Go's _simple_type.

Test package execution coverage

• Execute 13 previously omitted test packages in continuous integration.
• Check package assignments against the workflow and use the same plan for race execution.
• Reject new test packages without an execution lane.

Compact parser maintenance

• Move scheduler memory accounting and stop checks into a dedicated source file.
• Share symbol metadata projection between recovery and selected-store materialization.
• Include retained recovery-memo capacity in scheduler memory limits, including after reset.
• Document source ownership and the checks required for new retained state.

Recovery symbol allocation

• Reuse one symbol table for recovery per compact parse across all language grammars.
• Count the table in the memory budget and discard it when the scheduler resets.
• Reduce allocated bytes by 81.58 percent on the real Go query compiler fixture. Parse timing shows no significant change.

Recovery memo pressure

• Compute leaf error costs and visible counts without occupying recovery memo entries.
• Preserve error-region costs, missing-token costs, cache limits, and cleanup.

Compact incremental allocation

• Grow borrowed incremental arenas as needed instead of reserving capacity for the complete source.
• Preserve full-parse reservation, memory limits, and cleanup after a compact decline.

Incremental profile accounting

• Count discarded retry work without adding its reuse coverage to the selected result.
• Preserve fallback work, phase times, and maximum resource counts in incremental profiles.
• Keep reuse status and parse boundaries tied to the returned tree.

Standalone grammar packages

• Add generated standalone packages for all 206 blob-backed grammars without build tags.
• Share scanners, decoder repairs, caches, and certified profiles through grammars/runtime.
• Remove the aggregate catalog dependency from the native Lean package.

Generated source ownership

• Standardize generated Go markers and name each generator command as the
  owner. Continuous integration now rejects unknown owners, malformed markers,
  and generated file names without markers.
• Move the grammargen marker before the package clause. Go tools can now
  identify emitted grammar source as generated code.

C parity program, round one: query semantics

• Resolve node types in query patterns the way the C query compiler does:
  only a visible or supertype named symbol is a node type. A hidden rule
  name or an anonymous token in a node pattern is now a compile error, as in
  C. The inferred tags queries use the same lookup, so a grammar whose
  call is a keyword no longer receives a call pattern.
• Compile a supertype node pattern such as (expression) into a wildcard
  step that requires the supertype among the node's hidden ancestors, and
  support the super/sub form with the C subtype check. Nodes record the
  hidden supertype wrappers that reduction elided in a parallel arena table
  (Node stays 104 bytes); the record survives final tree compaction and
  incremental clones.
• Port the C wildcard-root rule: a pattern whose root is a wildcard (a
  supertype counts) and whose first child is a concrete node type never
  tests the root. (expression (identifier) @i) matches every identifier
  whose parent is not an ERROR node, as it does in C.
• Wildcard steps never match ERROR nodes, and a top-level bare _ pattern
  compiles.
• Share one Lua-pattern compiler between production queries and both C query comparisons.
  The outline comparison now evaluates #lua-match? predicates instead of rejecting them.
• TestParityQuerySemantics runs 103 query cases on both engines: 101
  agree, 2 carry a named divergence (an aliased subtype in the grammargen
  supertype map, and a hidden wrapper lost inside a compact error region).
  Highlight parity holds on 204 of 206 languages with no tolerance entry;
  hare and luau, the last two tolerated languages, now match C.
• TestParitySupertypeMap compares every grammar's ABI 15 supertype map
  with the C runtime: 40 languages agree, 29 diverge. The board is
  informational until the grammargen map is rebuilt.

C parity program, round two: recovery

• Add TestParityRecoveryBoard: 78 malformed sources in eight languages
  parsed on the C oracle and on every Go route, compared node by node. The
  default route agrees on 36 (29 before this round); with the C recovery
  port forced on for JavaScript, 46.
• An absorbed leaf inside an ERROR region carries no error bit, as in C,
  where only a missing leaf has an error cost. The region proof that used
  to decide when a leaf could stay clean is gone.
• The A0 dispatcher census receipts for cobol and wgsl, and the cooklang
  witness digests, now pin the trees without leaf error bits. The cobol
  MBANK30.cpy fixture matches the C oracle exactly.
• Keyword capture follows ts_parser__lex: a keyword stays a keyword when
  the parse state has an action for it or reserves it; otherwise the lexer
  returns the word token. The reserved-word rule was inverted before.
• A missing leaf takes an inherited field, as a relevant child does in C.
• cpp, html, javascript, and julia stay on the legacy recovery path behind
  measured witnesses recorded in docs/c-parity-boards.md.
• An accepted GLR version stays out of the stack merge, as C removes it
  from the version pool, so a recovery fork created at end of input still
  competes as its own tree. An ERROR node keeps the fields a hidden child
  gave its spliced children. The recovery board moves to 36 of 78.
• The incremental invariant gate records its first two entries: python
  setup.py byte 1241 (delete and replace) parses without an error bit on
  both routes while C reports an ERROR, and the fresh and incremental
  parses keep a different number of GLR stacks after the site. The C
  keyword rule exposed the site; the divergence itself is older.
• Retire the Python interpolation compatibility pass after native clean-tie
  election reaches parity with the pinned C parser.
• Keep raw-shape ordering for forest-local alternatives when compact primary
  derivation selection is certified. Python f-string splats now match locked C
  on the forest route.

Compact core cost, round two (issue #​454)

• Fuse the top-down parse-state replay into the postorder materialization
  visit. The visit computes each subtree's pre-goto and parse state at push
  time with the same transition rules, so the tree needs no second
  full-derivation pass and no arena-length replay tables.
  TestCompactFusedReplayMatchesTopDownReplay proves the states equal the
  separate replay on every subtree.
• Remove the dead tokenCell election record and its five save-and-restore
  sites, read the reuse-dependency subtree count and head path count through
  narrow accessors instead of Core.Stats, and build the election record in
  place.
• Stop copying large records on the hot path: headers, reduction outputs,
  pop paths, boundary outputs, and canonical groups are read through
  pointers; the
  direct-append condense reads the predecessor it already resolved instead of
  validating a synthetic link and resolving it again; a zero stored cost no
  longer republishes a fresh node's lineage.
• Validate link records at node publication, including copied adjacencies.
  Single-link pop enumeration can trust immutable published records.
  The relex payload scratch no longer clears its whole buffer on
  every election, the head owner record runs without a closure per dispatch,
  and a single fresh reduction output updates its header in place.
• Earlier exploratory measurements predate the correctness review and
  benchmark lifetime fixes. They do not establish current performance gains.
  The route decision record retains them as historical measurements.
• Extract the accepted-tree visit into compactMaterializer, a struct the
  scheduler can drive as well as the postorder pass. The postorder pass
  now fills one scratch view in place and visits it through a pointer, and
  it can skip subtrees that already own a public node
  (VisitMaterializationPostorderPrebuilt). The extraction changes no
  tree and no work count.
• Add the eager materialization lane (GTS_COMPACT_EAGER=1). After each
  single-header shift and each in-place reduction the scheduler builds the
  new subtree's public node at once, and it builds the subtrees a
  multi-header phase left pending as soon as a single header consumes them.
  On every Go witness the lane builds the whole tree before acceptance and
  publishes the same tree, the same replay stamps, and the same work as the
  postorder pass (TestCompactEagerMaterializationMatchesPostorder). The
  lane stays off by default: on the Go 137 KiB witness it costs about ten
  percent more wall time, because construction interleaved with dispatch
  loses the locality of the batch pass while the compact core still writes
  every record. The lane is the construction half of the single-head kernel,
  which will stop writing compact records for subtrees that already own a
  public node.
• Skip the canonical-boundary probe when a single header holds a node the
  dispatch just published: a fresh node is the latest node of its phase
  identity, so the probe would return the head the header already holds.
  The generic shift, the in-place reduction, and the corridor direct shift
  all take the skip when the header sits outside recovery isolation with no
  pending freshness; the skip records the barrier, the header peak, and the
  verifier binding, so every work vector and receipt stays identical. Parents take their span
  from the point index only when their visible children do not tile the
  record, and a reduction sums its pop payload work once.
• Turn the C4 bytecode corridor on by default (stage 3 of
  spec.c4-bytecode-isa.v1). The 137 KiB full-parse comparison is faster on
  14 of 15 grammars. A JavaScript recovery mutation once changed the C tree
  with the lane on; the lane now stays off while a version-owned lexer
  request is live, and the evidence for the default is: the runtime
  equivalence test keeps every work count and digest equal; the exhaustive
  curated structural parity suite (fresh, incremental, no-error) passes on
  every grammar with the lane on; the pinned-oracle T3 recovery adjudication
  in the harness container matches C on every html and JavaScript witness
  with the lane on; and the JavaScript recovery mutation differentials pass
  in both modes. GTS_C4_CORRIDOR=0 turns the lane off.
• Keep version-owned lexer requests on the generic dispatch path. The
  corridor reads a shared token and cannot publish an owned request.
• Preserve separate canonicalization output buffers for single headers.
  Reusing the input slice changed earlier snapshots and broke rollback isolation.
• Answer point lookups from the line of the previous answer or the next
  line before the hashed cache and the binary search: materialization asks
  for points in source order. Skip the scanner-provenance search for a
  terminal that cannot carry an entry, and the skipped-prefix search when
  no prefix was recorded. Together about 3 percent on the Go 137 KiB
  witness.

Production engine fixes kept until retirement (issue #​454)

The compact route stays the default fresh full-parse route. The owner's
direction is to retire the production engine once the compact core
outperforms it; until then production still serves incremental, injection,
included-range, and fallback parses, so these fixes stay.

• Isolate parser scratch lifetimes across parses. A pooled scratch kept the
  transient parent and child slabs of the largest earlier parse, up to 512K
  elements, and billed them to every later parse in the process: a 4 KiB
  parse after a 315 KiB parse reported 35 MB of inherited scratch. Each parse
  now drops inherited transient slabs above four times its own initial arena
  estimate before it starts. A new small-large-small test guards the bound
  through the new ParseRuntime.TransientScratchBytesAllocated counter.
• Shrink Token from 80 to 64 bytes. The five unexported provenance bits
  pack into one flag byte, and the stack position behind a synthetic missing
  token moves to a parser-owned anchor table that the token indexes. Tokens
  are copied by value on every election and dispatch, so the size shows up
  directly as copy cost on both routes. The public fields are unchanged.
• Bound reuse-hostile incremental parses. An old-tree reuse parse that has
  built four times the larger of the old tree's nodes and the fresh-parse
  arena estimate while reusing under one eighth of the source now stops with
  ParseStopReuseBudget, and the parser runs one plain full parse, the same
  fail-closed retry the memory budget uses. The issue #​454 C single-byte
  delete built 3.2 million nodes before the memory budget stopped it; it now
  stops near 370 thousand and returns the fresh-parse tree. The profile names
  the retry incremental_parse_reuse_budget_full_retry.

Compact route repair (issue #​454)

• Repair the three regressions on the compact candidate route that issue
  #​454 measured on
  137 KiB editor fixtures.
• Compact error recovery scales linearly. The recovery cost memo grew to the
  exact size on every store and was reallocated on every call, so a fresh
  parse of a 16 KiB Go file with one syntax error took 4.9 seconds. The memo
  now grows geometrically and lives for the whole parse. The same parse takes
  49 milliseconds, and the 137 KiB single-byte delete completes in 178
  milliseconds instead of never.
• Compact-materialized old trees reuse top-level siblings under the same
  compatible-goto contract as production trees. TOML insert reuse returns
  from 62 percent to 100 percent, and TypeScript from 54 percent to 98 percent.
• A synthesized root no longer disables reuse for the whole tree. INI files
  that end in a blank line return from 0 percent to 97 percent reuse. The
  unsupported-reuse reason now names the clause that failed.
• The compact incremental attempt declines after eight unauthenticated
  in-scope candidates or 32 KiB past the edit with zero reuse, so INI and
  JSON no longer pay a discarded whole-file compact parse per keystroke.
• The compact scheduler checks its current footprint at every memory-budget
  poll. A cached small footprint did not detect subsequent storage growth.
  Regression tests cover both the memory budget and the hard ceiling.
• The compact scheduler skips avoidable per-token work: the checkpoint
  interner compares against the last interned
  record before hashing, the relex probe authenticates its payload by byte
  comparison instead of SHA-256, and the materialization walk passes records
  by pointer. Earlier performance measurements predate the review fixes.
  Run randomized comparisons before reporting gains for the corrected code.
• Halt a production GLR stack at a no-action point when a sibling stack
  accepts the lookahead, before the previous-shift recovery runs. Pull
  request #​709 added a
  per-stack re-lex that kept the constructor-specifier fork of
  static inline void f(int *v) {} alive, and the recovered fork won
  selection with an ERROR node. Five C++ witnesses now match the compact route.
• Add cmd/issue454bench, which reproduces the downstream measurements on
  synthetic fixtures with an optional CPU profile.
• Serve a mid-file transient-error keystroke on a compact old tree with
  production incremental reuse, the v0.48.1 mechanism, when the compact
  borrow attempt declines at recovery. The fresh compact recovery route never
  produced those trees; it declined after a whole-file pass. Edits within 256
  bytes of end of file keep the compact recovery route. Go single-byte deletes
  drop from 178 to 15 milliseconds at 137 KiB, and every measured tree equals
  the fresh default-route parse except two pre-existing divergences that the
  new parity gate documents.
• Skip the fail-closed whole-file reparse after an incremental parse whose
  errors sit inside top-level items covering at most a quarter of the source.
  Pull request #​613's wide-stack condition fired on TypeScript's ordinary GLR
  ambiguity, so a single-byte delete at 137 KiB cost 296 milliseconds against
  78 at v0.48.1; it now costs 75. Degenerate results still retry.
• Decline an unpublishable compact recovery when its region commits instead
  of after a whole-file pass. A fresh compact parse of a 137 KiB Go file with
  a mid-file error drops from about 424 to 275 milliseconds; the production
  parse alone costs 240.
• Shrink Token from 88 to 80 bytes, memoize the scanner identity
  fingerprint per parse, and compute per-election checkpoint receipt digests
  only under full receipts. Scala and CMake compact full parses gain about
  another 12 percent.
• Cut the production engine's drift since v0.48.1, which both routes
  inherit. The token source passes tokens by pointer through its per-token
  helper chain instead of copying 80 bytes about ten times per token, both
  lexers decode the frontier rune only for non-ASCII bytes, the contextual
  close-angle probe checks the token bytes before symbol names, and the
  external scanner failure-mode probes are answered once per language.
  Production full parses of 137 KiB fixtures move from 1.1 to 1.4 times
  v0.48.1 to 1.06 to 1.19 times, with Rust at 1.33. The report attributes
  the remaining gap and records the compact route's graduation status.
• Remove four avoidable per-token costs from the compact scheduler: the
  cap-pressure poll reads the node count without validating the head, the
  per-state relex probe caches the scanner contract and identity and uses
  scheduler-owned snapshot scratch, the election reads the cached checkpoint
  identity instead of asking the order adapter, and the reuse-proof
  invalidation takes the lineage record by pointer. Go compact full parses
  gain 7 percent; the other grammars are within 2 percent.

Compact parser correctness

• Authenticate terminal aliases at ordinary grammar reductions during recovery. Preserve separate rules for synthetic ERROR reductions and retain span coverage checks.
• Preserve inherited alternative history when a reduction joins an active sibling.
• Retain convergence and resurrection restrictions after adoption. Preserve existing blended-history rejection checks.
• Parse foo<A00>(2); through compact without fallback, with exact locked-C tree parity.
  Malformed TypeScript recovery remains unfinished.
• Preserve numeric-edit reuse for newly admitted TypeScript trees after authenticating lexer dependencies and scanner equivalence.

Incremental correctness

• Restore bounded token-invariant reuse after authenticating earlier lexical
  dependencies and the edited token. Unknown coverage requires reparsing.
• Retain examined-byte coverage through failed scans, rollback, and accepted
  tree ownership. Include UTF-8 continuation bytes beyond token boundaries.
• Add optional scanner byte-equivalence declarations. These declarations do
  not authorize general subtree reuse or replace scanner checkpoints.
• Check repeated edits against locked C trees for Go, CSS, SCSS, TypeScript,
  and Julia. Keep malformed and unsupported cases on their existing fallback paths.

Twenty paired benchmark samples compare this change with v0.52.0.
Generated Go single-byte edits improve from 3,033.2 to 182.1 microseconds.
Allocations decrease from 95 to 3 per edit. Full parsing regresses 1.68 percent.
See pull request #​1093 for the restoration and its validation.
These changes do not complete compact parser graduation or retire the legacy parser.

---

Configuration

📅 Schedule: (UTC)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR was generated by Mend Renovate. View the repository job log.

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 2321ac02-e38f-44cc-8a27-6701d8d04a3e

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.15%. Comparing base (c6f4b67) to head (d8fe54a).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #363      +/-   ##
==========================================
- Coverage   88.17%   88.15%   -0.02%     
==========================================
  Files         362      362              
  Lines       34743    34743              
==========================================
- Hits        30634    30629       -5     
- Misses       4104     4109       +5     
  Partials        5        5              

Flag	Coverage Δ
unittests	88.15% <ø> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.