diff --git a/src/parse-anplusb.ts b/src/parse-anplusb.ts
index f160a0b..2120d2a 100644
--- a/src/parse-anplusb.ts
+++ b/src/parse-anplusb.ts
@@ -4,7 +4,7 @@
 // Much inspiration taken from CSSTree (Roman Dvornov) - MIT License
 // https://github.com/csstree/csstree/blob/56afb6dd761149099cd3cdfb0a38e15e8cc0a71a/lib/syntax/node/AnPlusB.js#L106-L271
 
-import { Lexer } from './tokenize'
+import { Lexer, type LexerPosition } from './tokenize'
 import { NTH_SELECTOR, CSSDataArena } from './arena'
 import {
 	TOKEN_IDENT,
@@ -23,6 +23,8 @@ export class ANplusBParser {
 	private arena: CSSDataArena
 	private source: string
 	private expr_end: number
+	// Reused "+n" lookahead snapshot (single use per parse_anplusb call, never nested)
+	private plus_n_lookahead_position: LexerPosition = new Lexer('').save_position()
 
 	constructor(arena: CSSDataArena, source: string) {
 		this.arena = arena
@@ -152,7 +154,7 @@ export class ANplusBParser {
 			this.source.charCodeAt(this.lexer.token_start) === CHAR_PLUS
 		) {
 			// Look ahead for 'n'
-			const saved = this.lexer.save_position()
+			const saved = this.lexer.save_position_into(this.plus_n_lookahead_position)
 			this.lexer.next_token_fast(true)
 
 			if ((this.lexer.token_type as TokenType) === TOKEN_IDENT) {
diff --git a/src/parse-atrule-prelude.ts b/src/parse-atrule-prelude.ts
index 194a6ad..a986de0 100644
--- a/src/parse-atrule-prelude.ts
+++ b/src/parse-atrule-prelude.ts
@@ -1,5 +1,5 @@
 // At-Rule Prelude Parser - Parses at-rule preludes into structured AST nodes
-import { Lexer } from './tokenize'
+import { Lexer, type LexerPosition } from './tokenize'
 import {
 	CSSDataArena,
 	MEDIA_QUERY,
@@ -50,6 +50,11 @@ export class AtRulePreludeParser {
 	// Shared media-feature/supports-condition/style()-selector() parsing, also used by ValueNodeParser
 	// for if()'s condition functions — see parse-condition.ts for why value_node_parser is injected
 	private condition_parser: ConditionParser
+	// Reused lookahead snapshot shared by all save/restore sites in this class — none of
+	// them stay live across a call that could itself need this slot (verified: none here
+	// recurse back into this class, and the two @import helpers resolve their save before
+	// delegating further parsing elsewhere).
+	private lookahead_position: LexerPosition = new Lexer('').save_position()
 
 	constructor(arena: CSSDataArena, source: string) {
 		this.arena = arena
@@ -132,7 +137,7 @@ export class AtRulePreludeParser {
 
 			// Skip comma separator
 			this.skip_whitespace()
-			const saved = this.lexer.save_position()
+			const saved = this.lexer.save_position_into(this.lookahead_position)
 			this.next_token()
 			if (this.lexer.token_type !== TOKEN_COMMA) {
 				// Not a comma, restore position
@@ -203,7 +208,7 @@ export class AtRulePreludeParser {
 		let last_component = 0
 
 		// Leading modifier (only/not) — emit as PRELUDE_OPERATOR like the and/or/not combinators below
-		const saved_token_start = this.lexer.save_position()
+		const saved_token_start = this.lexer.save_position_into(this.lookahead_position)
 		this.next_token()
 
 		if (this.lexer.token_type === TOKEN_IDENT) {
@@ -390,7 +395,7 @@ export class AtRulePreludeParser {
 
 				// Glue on '.' ident segments immediately following, with no gaps.
 				while (this.lexer.pos < this.prelude_end) {
-					let saved = this.lexer.save_position()
+					let saved = this.lexer.save_position_into(this.lookahead_position)
 
 					let dot_token_type = this.next_token()
 					if (
@@ -556,7 +561,7 @@ export class AtRulePreludeParser {
 	// Parse import layer: layer or layer(name)
 	private parse_import_layer(): number | null {
 		// Peek at next token
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.lookahead_position)
 
 		this.next_token()
 
@@ -609,7 +614,7 @@ export class AtRulePreludeParser {
 	// Parse import supports: supports(condition)
 	private parse_import_supports(): number | null {
 		// Peek at next token
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.lookahead_position)
 
 		this.next_token()
 
@@ -662,7 +667,7 @@ export class AtRulePreludeParser {
 
 	// Helper: Peek at next token type without consuming
 	private peek_token_type(): number {
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.lookahead_position)
 
 		this.next_token()
 		let type = this.lexer.token_type
@@ -690,7 +695,7 @@ export class AtRulePreludeParser {
 		if (this.lexer.pos >= this.prelude_end) return []
 
 		// Peek: if it's an IDENT it's an optional prefix, otherwise let parse_import_url() take it
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.lookahead_position)
 		this.next_token()
 
 		if (this.lexer.token_type === TOKEN_IDENT) {
diff --git a/src/parse-condition.ts b/src/parse-condition.ts
index 3f1eda6..668bafe 100644
--- a/src/parse-condition.ts
+++ b/src/parse-condition.ts
@@ -8,7 +8,7 @@
 // own (a plain `import { ValueNodeParser } from './value-node-parser'` here would be
 // circular, since value-node-parser.ts imports this file). The type-only import below is
 // erased at build time and creates no runtime dependency in either direction.
-import { Lexer } from './tokenize'
+import { Lexer, type LexerPosition } from './tokenize'
 import {
 	CSSDataArena,
 	MEDIA_FEATURE,
@@ -61,6 +61,9 @@ export class ConditionParser {
 	// Deep-parses selector()'s argument into a real SelectorList. No cycle risk (SelectorParser
 	// doesn't depend on ValueNodeParser/AtRulePreludeParser), so this is safely owned here.
 	private selector_parser: SelectorParser
+	// Reused peek snapshot for parse_feature_range's token-type probe (never nested — resolved
+	// within the same loop iteration, before the next save).
+	private feature_range_peek_position: LexerPosition = new Lexer('').save_position()
 
 	constructor(arena: CSSDataArena, source: string, value_node_parser: ValueNodeParser) {
 		this.arena = arena
@@ -259,7 +262,7 @@ export class ConditionParser {
 				last_child = op
 			} else {
 				// Value or feature name
-				let saved = this.lexer.save_position()
+				let saved = this.lexer.save_position_into(this.feature_range_peek_position)
 				this.lexer.pos = pos
 				this.next_token()
 
diff --git a/src/parse-declaration.ts b/src/parse-declaration.ts
index a8eee30..aec4306 100644
--- a/src/parse-declaration.ts
+++ b/src/parse-declaration.ts
@@ -39,6 +39,9 @@ export class DeclarationParser {
 	private value_parser: ValueParser | null
 	// Reused lookahead snapshot (never nested) to avoid an allocation per declaration
 	private lookahead_position: LexerPosition = new Lexer('').save_position()
+	// Separate scratch slot: the browser-hack delimiter-prefix snapshot (below) stays live
+	// at the same time as lookahead_position, so it can't reuse that field.
+	private delim_prefix_position: LexerPosition = new Lexer('').save_position()
 
 	constructor(arena: CSSDataArena, source: string, parse_values: boolean = true) {
 		this.arena = arena
@@ -111,7 +114,7 @@ export class DeclarationParser {
 			lexer.token_type === TOKEN_COLON
 		) {
 			// Browser hacks can use various token types as prefixes (e.g., * hack)
-			const delim_saved = lexer.save_position()
+			const delim_saved = lexer.save_position_into(this.delim_prefix_position)
 			browser_hack_start = lexer.token_start
 			browser_hack_line = lexer.token_line
 			browser_hack_column = lexer.token_column
diff --git a/src/parse-selector.ts b/src/parse-selector.ts
index 6fcd62b..97186fd 100644
--- a/src/parse-selector.ts
+++ b/src/parse-selector.ts
@@ -1,5 +1,5 @@
 // Selector Parser - Parses CSS selectors into structured AST nodes
-import { Lexer } from './tokenize'
+import { Lexer, type LexerPosition } from './tokenize'
 import { CSSDataArena } from './arena'
 import {
 	SELECTOR,
@@ -71,6 +71,30 @@ export class SelectorParser {
 	private source: string
 	private selector_end: number
 	private anplusb_parser: ANplusBParser
+	// Reused lookahead snapshots for the save/restore sites that are never live across a
+	// recursive re-entry into the selector grammar (unlike parse_compound_selector's,
+	// parse_pseudo_function_after_colon's, and parse_nth_expression's own saves, which stay
+	// allocating for now — see the recursive :is()/:has()/:nth-child(of) call chain).
+	// Resolved entirely before parse_complex_selector's compound-selector loop runs.
+	private complex_selector_lookahead_position: LexerPosition = new Lexer('').save_position()
+	// Live only inside parse_namespace_local_part itself, never during its own call into
+	// anything else — but concurrently live with whichever of the two fields below called it,
+	// so it must stay distinct from both.
+	private namespace_local_part_lookahead_position: LexerPosition = new Lexer('').save_position()
+	// Shared by parse_type_or_namespace_selector and parse_universal_or_namespace_selector:
+	// alternative call paths (type selector vs universal selector) that are never both active,
+	// each held live across its own call into parse_namespace_local_part.
+	private namespace_selector_lookahead_position: LexerPosition = new Lexer('').save_position()
+	// try_parse_combinator: fully resolved within one call, no nested saves.
+	private combinator_lookahead_position: LexerPosition = new Lexer('').save_position()
+	// parse_class_selector: fully resolved within one call, no nested saves.
+	private class_selector_lookahead_position: LexerPosition = new Lexer('').save_position()
+	// parse_pseudo's two concurrent saves (`saved` persists longer, discarded rather than
+	// re-read on the path that recurses into parse_pseudo_function_after_colon).
+	private pseudo_lookahead_position: LexerPosition = new Lexer('').save_position()
+	private pseudo_double_colon_lookahead_position: LexerPosition = new Lexer('').save_position()
+	// parse_lang_identifiers: fully resolved within one call, no nested saves.
+	private lang_identifiers_lookahead_position: LexerPosition = new Lexer('').save_position()
 
 	constructor(arena: CSSDataArena, source: string) {
 		this.arena = arena
@@ -192,7 +216,7 @@ export class SelectorParser {
 
 		// Check for leading combinator (relative selector) if allowed
 		if (allow_relative && this.lexer.pos < this.selector_end) {
-			const saved = this.lexer.save_position()
+			const saved = this.lexer.save_position_into(this.complex_selector_lookahead_position)
 
 			this.lexer.next_token_fast(false)
 			let token_type = this.lexer.token_type
@@ -357,7 +381,7 @@ export class SelectorParser {
 		namespace_start: number,
 		namespace_prefix_length: number,
 	): number | null {
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.namespace_local_part_lookahead_position)
 		this.lexer.next_token_fast(false)
 
 		let node_type: number
@@ -404,7 +428,7 @@ export class SelectorParser {
 		}
 
 		// Save position before skipping whitespace/comments
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.namespace_selector_lookahead_position)
 
 		// Skip whitespace and comments before checking for |
 		this.skip_whitespace()
@@ -429,7 +453,7 @@ export class SelectorParser {
 	// Called when we've seen a * DELIM token
 	private parse_universal_or_namespace_selector(start: number, end: number): number | null {
 		// Save position before skipping whitespace/comments
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.namespace_selector_lookahead_position)
 
 		// Skip whitespace and comments before checking for |
 		this.skip_whitespace()
@@ -459,7 +483,7 @@ export class SelectorParser {
 
 	// Parse combinator (>, +, ~, or descendant space)
 	private try_parse_combinator(): number | null {
-		const saved_whitespace_start = this.lexer.save_position()
+		const saved_whitespace_start = this.lexer.save_position_into(this.combinator_lookahead_position)
 		let has_whitespace = this.lexer.pos < this.selector_end
 
 		// Skip whitespace and comments
@@ -513,7 +537,7 @@ export class SelectorParser {
 	// Parse class selector (.classname)
 	private parse_class_selector(dot_pos: number): number | null {
 		// Save lexer state for potential restoration
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.class_selector_lookahead_position)
 
 		// Next token should be identifier
 		this.lexer.next_token_fast(false)
@@ -685,10 +709,10 @@ export class SelectorParser {
 	// Parse pseudo-class or pseudo-element (:hover, ::before)
 	private parse_pseudo(start: number): number | null {
 		// Save lexer state for potential restoration
-		const saved = this.lexer.save_position()
+		const saved = this.lexer.save_position_into(this.pseudo_lookahead_position)
 
 		// Save position before skipping whitespace/comments
-		const saved_ws = this.lexer.save_position()
+		const saved_ws = this.lexer.save_position_into(this.pseudo_double_colon_lookahead_position)
 
 		// Skip whitespace and comments before checking for second colon
 		this.skip_whitespace()
@@ -840,7 +864,7 @@ export class SelectorParser {
 	// Parse :lang() content - comma-separated language identifiers
 	// Accepts both quoted strings: :lang("en", "fr") and unquoted: :lang(en, fr)
 	private parse_lang_identifiers(start: number, end: number, parent_node: number): void {
-		const saved_position = this.lexer.save_position()
+		const saved_position = this.lexer.save_position_into(this.lang_identifiers_lookahead_position)
 		this.lexer.seek(start, this.lexer.line, this.lexer.column)
 
 		let saved_selector_end = this.selector_end
diff --git a/src/value-node-parser.ts b/src/value-node-parser.ts
index 4e72eb7..64fa177 100644
--- a/src/value-node-parser.ts
+++ b/src/value-node-parser.ts
@@ -1,7 +1,7 @@
 // Value Node Parser - shared recursive value-token parsing, used by both ValueParser
 // (declaration values) and AtRulePreludeParser (feature values), so calc()/env()/var()
 // get the same structured tree everywhere instead of an opaque text span.
-import { Lexer } from './tokenize'
+import { Lexer, type LexerPosition } from './tokenize'
 import {
 	CSSDataArena,
 	IDENTIFIER,
@@ -66,6 +66,13 @@ export class ValueNodeParser {
 	// try to build another ValueNodeParser to inject, forever. The helper's own condition_parser is
 	// simply never accessed (it's only ever used via parse_chain), so the recursion stops there.
 	private _condition_parser: ConditionParser | null = null
+	// Reused peek snapshots. Two distinct fields: parse_function_node's url()/src() peek and
+	// parse_if_condition's and/or/not continuation peek are each resolved before anything else
+	// in this class could save again, so neither is ever live at the same time as the other —
+	// but parse_value_nodes_in_range's own save (elsewhere) CAN be live across a recursive
+	// descent into this class, so it deliberately does not share either of these.
+	private url_peek_position: LexerPosition = new Lexer('').save_position()
+	private if_condition_peek_position: LexerPosition = new Lexer('').save_position()
 
 	private get condition_parser(): ConditionParser {
 		if (this._condition_parser === null) {
@@ -266,7 +273,7 @@ export class ValueNodeParser {
 		if (is_url || str_equals_range(this.source, start, name_end, 'src')) {
 			// Peek at the next token to see if it's a string
 			// If it's a string, parse normally. Otherwise, skip parsing children.
-			let save_pos = this.lexer.save_position()
+			let save_pos = this.lexer.save_position_into(this.url_peek_position)
 			this.lexer.next_token_fast(false)
 
 			// Skip whitespace
@@ -593,7 +600,7 @@ export class ValueNodeParser {
 			// condition; anything else ends the condition here.
 			if (tt !== TOKEN_FUNCTION && !is_operator) break
 
-			let saved = this.lexer.save_position()
+			let saved = this.lexer.save_position_into(this.if_condition_peek_position)
 			let next_tt = this.next_significant_token()
 			if (next_tt === TOKEN_FUNCTION) {
 				compound = true