rspec-tracer 1.2.2 → 2.0.0.pre.1

data/lib/rspec_tracer/storage/json_backend.rb

@@ -0,0 +1,866 @@
+# frozen_string_literal: true
+
+require 'digest/md5'
+require 'fileutils'
+require 'json'
+require 'set'
+require 'time' # Time#iso8601 for last_run.json timestamp
+
+require_relative 'backend'
+require_relative 'lazy_snapshot'
+require_relative 'schema'
+require_relative 'serializer/json'
+require_relative 'serializer/msgpack'
+require_relative 'snapshot'
+
+module RSpecTracer
+  # Internal Storage — see {RSpecTracer} for the user-facing surface.
+  # @api private
+  module Storage
+    # JSON-on-disk storage backend. 1.x shipped this layout without a
+    # formal contract; 2.0 treats the FILENAMES list below as the
+    # authoritative user-facing surface.
+    #
+    # External tooling (CI cache keys, debug scripts, report
+    # renderers) may reference these exact filenames, so additions or
+    # removals are breaking changes. The shared-examples contract in
+    # `spec/contracts/storage_backend.rb` enforces the list.
+    #
+    # Commit point: `last_run.json` is written last via tmp + rename.
+    # If any of the 11 per-run files fails to write, `last_run.json`
+    # stays pointed at the previous successful run and the partially-
+    # written run-id directory is orphaned (harmless; `clear!` reaps
+    # it). Readers that see `last_run.json` therefore see a complete
+    # snapshot.
+    #
+    # Concurrency: an exclusive flock on a sentinel file
+    # (`.rspec_tracer.lock` under cache_path) serializes writers.
+    # Readers do not take the lock - `last_run.json`'s atomic rename
+    # is their consistency model.
+    #
+    # Corruption policy: `load_graph` never raises. Missing files,
+    # malformed JSON, wrong schema, binary-garbage input all yield
+    # `nil` + an info log. This is the invariant the fuzz spec
+    # asserts across 1000 iterations.
+    #
+    # Encoding: every read and write passes `encoding: 'UTF-8'`.
+    # Fixes the `Encoding::InvalidByteSequenceError` that bit the
+    # dogfood path when an example title contained a non-ASCII byte on
+    # a US-ASCII-defaulted filesystem.
+    # rubocop:disable Metrics/ClassLength
+    class JsonBackend
+      # On-disk filenames under the default `:json` serializer. This
+      # is the user-facing surface documented in
+      # USER_FACING_SURFACE.md section 6 - external tooling that walks
+      # `rspec_tracer_cache/` relies on exactly these names.
+      # The `:msgpack` serializer substitutes `.msgpack.gz` for the
+      # `.json` suffix (one file per field on disk); the file stems
+      # and per-field semantics do not change.
+      # boot_set.json lands at the end of the list - additive w.r.t.
+      # 1.x and v2 readers that walked this enumeration. It carries
+      # the project's transitive boot-load set (schema_version 3).
+      # wsi_snapshot.json persists the WholeSuiteInvalidators
+      # digest_snapshot so warm runs can tell whether Gemfile.lock /
+      # .ruby-version / .rspec-tracer / tracer-gem identity changed
+      # since the previous run. Without it, warm runs always saw a
+      # nil previous and treated every run as a cold first run.
+      # Missing file deserializes to `{}` so older caches still load -
+      # the fallback path fires one full re-run (safe).
+      # env_snapshot.json persists the `Tracker::EnvSnapshot` digest
+      # map for env-var values the per-example `tracks: { env: ... }`
+      # DSL declares. Same missing-coerces-to-`{}` fallback as
+      # wsi_snapshot - no schema bump.
+      # env_dependency.json persists the per-example tracked-env
+      # attribution map that reporters need for the Examples Dependency
+      # report. Missing file coerces to `{}`; older caches load
+      # without a cold re-run.
+      FILENAMES = %w[
+        all_examples.json
+        duplicate_examples.json
+        interrupted_examples.json
+        flaky_examples.json
+        failed_examples.json
+        pending_examples.json
+        skipped_examples.json
+        all_files.json
+        dependency.json
+        reverse_dependency.json
+        examples_coverage.json
+        boot_set.json
+        wsi_snapshot.json
+        env_snapshot.json
+        env_dependency.json
+        cache_hit_reason.json
+      ].freeze
+
+      # Internal constant.
+      # @api private
+      LAST_RUN_FILENAME = 'last_run.json'
+      # Internal constant.
+      # @api private
+      LOCK_FILENAME = '.rspec_tracer.lock'
+      # Internal constant.
+      # @api private
+      ENCODING = 'UTF-8'
+
+      # Known snapshot field symbols. Derived directly from FIELD_KINDS
+      # below (the write-side and read-side shape tables both enumerate
+      # the same set, so a divergence would already blow up write
+      # paths). Kept as an Array of Symbol so `#read_field` can dispatch
+      # without constructing a per-serializer filename table; the
+      # filename is computed as "#{field}.#{@serializer.extension}".
+      FIELD_NAMES = %i[
+        all_examples
+        duplicate_examples
+        interrupted_examples
+        flaky_examples
+        failed_examples
+        pending_examples
+        skipped_examples
+        all_files
+        dependency
+        reverse_dependency
+        examples_coverage
+        boot_set
+        wsi_snapshot
+        env_snapshot
+        env_dependency
+        cache_hit_reason
+      ].freeze
+
+      # Binds a backend + run directory so `LazySnapshot` readers
+      # call exactly one public entry point (`backend.read_field`).
+      # Keeping this as a nested class (not a Proc) so mutant can
+      # introspect the reader contract.
+      class FieldReader
+        # Internal method on the tracer pipeline.
+        # @api private
+        def initialize(backend:, dir:)
+          @backend = backend
+          @dir = dir
+        end
+
+        # Internal method on the tracer pipeline.
+        # @api private
+        def read(field)
+          @backend.read_field(@dir, field)
+        end
+      end
+
+      # Write-side field groups. Each group dispatches to one
+      # serializer (Hash pass-through, Set->sorted Array, or the
+      # Hash[id => Set<path>] -> Hash[id => Array<path>] flavor
+      # shared by dependency + reverse_dependency). Kept data-driven
+      # so a schema_version bump adds one entry instead of a new
+      # branch. Read-side uses FIELD_KINDS below.
+      ID_SET_FIELDS = %w[
+        interrupted_examples flaky_examples failed_examples pending_examples skipped_examples
+      ].freeze
+      # Internal constant.
+      # @api private
+      HASH_FIELDS = %w[
+        all_examples duplicate_examples all_files examples_coverage
+        boot_set wsi_snapshot env_snapshot env_dependency cache_hit_reason
+      ].freeze
+      # Internal constant.
+      # @api private
+      DEPENDENCY_FIELDS = %w[dependency reverse_dependency].freeze
+
+      # Read-side field -> deserializer-kind map. Drives
+      # `decode_field` so the lazy reader looks up one shape
+      # per field instead of spelling out a case/when that
+      # has to stay in sync with FILENAMES. `:symbolized` =
+      # Hash whose inner Hash values get symbolized keys
+      # (1.x's all_examples / all_files convention);
+      # `:dupe_examples` = same but Array-of-inner-Hash;
+      # `:id_set` = Array on disk -> Set in memory;
+      # `:dependency` = Hash[id => Array] -> Hash[id => Set];
+      # `:plain_hash` = pass-through (examples_coverage, the
+      # digest maps, env_dependency).
+      FIELD_KINDS = {
+        all_examples: :symbolized,
+        all_files: :symbolized,
+        duplicate_examples: :dupe_examples,
+        interrupted_examples: :id_set,
+        flaky_examples: :id_set,
+        failed_examples: :id_set,
+        pending_examples: :id_set,
+        skipped_examples: :id_set,
+        dependency: :dependency,
+        reverse_dependency: :dependency,
+        examples_coverage: :plain_hash,
+        boot_set: :plain_hash,
+        wsi_snapshot: :plain_hash,
+        env_snapshot: :plain_hash,
+        env_dependency: :plain_hash,
+        cache_hit_reason: :plain_hash
+      }.freeze
+
+      # Internal attribute.
+      # @api private
+      attr_reader :cache_path, :serializer, :serializer_name
+
+      # rubocop:disable Metrics/ParameterLists
+      def initialize(cache_path:, logger: nil, retention_local_count: nil,
+                     warn_per_file_mb: nil, warn_total_mb: nil, serializer: :json)
+        # rubocop:enable Metrics/ParameterLists
+        @cache_path = File.expand_path(cache_path)
+        @logger = logger
+        @retention_local_count = retention_local_count
+        @warn_per_file_mb = warn_per_file_mb
+        @warn_total_mb = warn_total_mb
+        @serializer = resolve_serializer(serializer)
+        @serializer_name = serializer
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def last_run_id
+        manifest = read_last_run_manifest
+        return nil unless manifest.is_a?(Hash)
+
+        run_id = manifest['run_id']
+        return nil if run_id.nil? || run_id.to_s.empty?
+
+        run_id
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def load_graph(schema_version:)
+        manifest = read_last_run_manifest
+        return nil unless manifest.is_a?(Hash)
+
+        stored = manifest['schema_version']
+        unless Schema.supported?(stored) && stored == schema_version
+          info("schema_version mismatch (stored=#{stored.inspect}, expected=#{schema_version}); cold run")
+          return nil
+        end
+
+        run_id = manifest['run_id']
+        return nil if run_id.nil? || run_id.to_s.empty?
+
+        dir = File.join(@cache_path, run_id)
+        return nil unless File.directory?(dir)
+
+        LazySnapshot.new(
+          schema_version: stored, run_id: run_id,
+          reader: FieldReader.new(backend: self, dir: dir)
+        )
+      rescue StandardError => e
+        info("failed to load cache: #{e.class}: #{e.message}; cold run")
+        nil
+      end
+
+      # Read and deserialize one per-run field. Public so
+      # `FieldReader` (constructed by `load_graph`) can dispatch.
+      # Missing file -> same default value the eager read previously
+      # produced (Set.new for ID-set fields, {} for hashes) -
+      # preserves the "malformed cache loads gracefully" contract.
+      #
+      # `deep_intern` runs before the decode so String dedup
+      # happens once per on-disk path / example_id regardless of
+      # how many times the value appears in the parsed tree.
+      # RAM win on large caches is the whole point of this method;
+      # see json_backend_spec.rb "string interning" for the
+      # measurable assertion.
+      def read_field(dir, field)
+        raise ArgumentError, "unknown snapshot field: #{field.inspect}" unless FIELD_KINDS.key?(field)
+
+        raw = read_run_file(dir, field_filename(field))
+        decode_field(field, deep_intern(raw))
+      end
+
+      # Per-serializer on-disk filename for a snapshot field.
+      # `:json` -> `all_examples.json`; `:msgpack` ->
+      # `all_examples.msgpack.gz`. Public so integration specs /
+      # reporters can resolve the expected path without reaching
+      # into @serializer.
+      def field_filename(field)
+        "#{field}.#{@serializer.extension}"
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def save_graph(snapshot, schema_version:)
+        raise ArgumentError, 'snapshot must not be nil' if snapshot.nil?
+
+        unless Schema.supported?(schema_version)
+          raise ArgumentError, "unsupported schema_version: #{schema_version.inspect}"
+        end
+
+        run_id = snapshot.run_id
+        raise ArgumentError, 'snapshot.run_id must be a non-empty string' if run_id.nil? || run_id.to_s.empty?
+
+        transactional_save do
+          dir = File.join(@cache_path, run_id)
+          FileUtils.mkdir_p(dir)
+          write_run_files(dir, snapshot)
+          write_last_run_atomic(schema_version: schema_version, run_id: run_id)
+        end
+
+        maybe_prune_after_save
+        maybe_warn_size_budget(run_id)
+        snapshot
+      end
+
+      # Retain the `keep` most-recently-modified run-id directories
+      # under cache_path and delete older ones. Always preserves the
+      # run-id that `last_run.json` points at (deleting it would make
+      # the next reader cold-run). Returns the count removed. Never
+      # raises - a prune failure is logged at warn level and treated
+      # as best-effort cleanup, same graceful-degradation contract
+      # the remote cache backends use.
+      #
+      # `keep` nil / non-positive -> no-op. Called automatically from
+      # `save_graph` when the backend was constructed with
+      # `retention_local_count:`; also exposed via `rake
+      # rspec_tracer:cache:gc` for one-off cleanup.
+      def prune_run_dirs!(keep:)
+        return 0 if keep.nil? || keep <= 0
+        return 0 unless File.directory?(@cache_path)
+
+        current = last_run_id
+        candidates = collect_run_dirs
+        return 0 if candidates.empty?
+
+        _keep, pruned = partition_dirs_to_prune(candidates, keep: keep, current: current)
+        pruned.each { |path| FileUtils.rm_rf(path) }
+        pruned.size
+      rescue StandardError => e
+        @logger&.warn("rspec-tracer cache gc: prune failed (#{e.class}: #{e.message})")
+        0
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def transactional_save(&block)
+        raise ArgumentError, 'block required' unless block
+
+        FileUtils.mkdir_p(@cache_path)
+        File.open(lock_path, File::RDWR | File::CREAT, 0o644) do |lock|
+          lock.flock(File::LOCK_EX)
+          yield
+        end
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def clear!
+        return unless File.directory?(@cache_path)
+
+        FileUtils.rm_rf(@cache_path)
+      end
+
+      # Merge per-worker snapshots (written to `peer_cache_paths`) into
+      # this backend's top-level cache and persist via `save_graph`.
+      # Read each peer via `load_graph` so schema + corruption policy
+      # (missing files yield nil, malformed JSON logs + returns nil)
+      # flows through the same path as a normal load.
+      #
+      # No peers / every peer nil -> no-op returns nil. Partial peers
+      # merge what's available; graceful degradation is the entire
+      # point of running this at at_exit time.
+      #
+      # `schema_version` is passed through so peers saved under a
+      # different schema version are rejected without side effects
+      # (same semantics as a warm run under a mismatched cache).
+      def merge_from_peers(peer_cache_paths, schema_version:)
+        peer_snapshots = peer_cache_paths.filter_map do |path|
+          self.class.new(cache_path: path, logger: @logger, serializer: @serializer_name)
+            .load_graph(schema_version: schema_version)
+        end
+
+        return nil if peer_snapshots.empty?
+
+        merged = Merger.call(peer_snapshots, schema_version: schema_version)
+        save_graph(merged, schema_version: schema_version)
+        merged
+      end
+
+      # Stateless snapshot union. parallel_tests partitions spec files
+      # across workers, so example IDs are disjoint in practice - the
+      # merge collision rules (first-wins for metadata, sum-of-ints for
+      # per-line coverage) only fire on collaborating workers that
+      # happened to observe the same input file.
+      module Merger
+        # Internal helper for the tracer pipeline.
+        # @api private
+        def self.call(snapshots, schema_version:)
+          state = empty_state
+          snapshots.each { |s| absorb(state, s) }
+
+          state[:reverse_dependency] = reverse_of(state[:dependency])
+          state[:run_id] = Digest::MD5.hexdigest(state[:all_examples].keys.sort.to_json)
+
+          Snapshot.new(
+            schema_version: schema_version,
+            run_id: state[:run_id],
+            all_examples: state[:all_examples],
+            duplicate_examples: state[:duplicate_examples],
+            interrupted_examples: state[:interrupted_examples],
+            flaky_examples: state[:flaky_examples],
+            failed_examples: state[:failed_examples],
+            pending_examples: state[:pending_examples],
+            skipped_examples: state[:skipped_examples],
+            all_files: state[:all_files],
+            dependency: state[:dependency],
+            reverse_dependency: state[:reverse_dependency],
+            examples_coverage: state[:examples_coverage],
+            boot_set: state[:boot_set],
+            wsi_snapshot: state[:wsi_snapshot],
+            env_snapshot: state[:env_snapshot],
+            env_dependency: state[:env_dependency],
+            cache_hit_reason: state[:cache_hit_reason]
+          )
+        end
+
+        # Internal helper for the tracer pipeline.
+        # @api private
+        def self.empty_state
+          {
+            all_examples: {},
+            duplicate_examples: Hash.new { |h, k| h[k] = [] },
+            interrupted_examples: Set.new,
+            flaky_examples: Set.new,
+            failed_examples: Set.new,
+            pending_examples: Set.new,
+            skipped_examples: Set.new,
+            all_files: {},
+            dependency: Hash.new { |h, k| h[k] = Set.new },
+            examples_coverage: {},
+            boot_set: {},
+            wsi_snapshot: {},
+            env_snapshot: {},
+            env_dependency: {},
+            cache_hit_reason: Hash.new(0)
+          }
+        end
+
+        # Union every field from one peer snapshot into the running
+        # state. Each field has a distinct combine rule (merge-first-wins,
+        # Set#merge, concat, or summing coverage strengths), so the
+        # branching is inherent to the shape. Decomposing per-field would
+        # scatter the merge contract.
+        # rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/PerceivedComplexity
+        def self.absorb(state, snapshot)
+          state[:all_examples].merge!(snapshot.all_examples || {}) { |_, v, _| v }
+          (snapshot.duplicate_examples || {}).each do |id, entries|
+            state[:duplicate_examples][id].concat(entries)
+          end
+          state[:interrupted_examples].merge(snapshot.interrupted_examples || Set.new)
+          state[:flaky_examples].merge(snapshot.flaky_examples || Set.new)
+          state[:failed_examples].merge(snapshot.failed_examples || Set.new)
+          state[:pending_examples].merge(snapshot.pending_examples || Set.new)
+          state[:skipped_examples].merge(snapshot.skipped_examples || Set.new)
+          state[:all_files].merge!(snapshot.all_files || {}) { |_, v, _| v }
+          (snapshot.dependency || {}).each do |id, paths|
+            state[:dependency][id].merge(paths)
+          end
+          merge_examples_coverage!(state[:examples_coverage], snapshot.examples_coverage || {})
+          state[:boot_set].merge!(snapshot.boot_set || {})
+          state[:wsi_snapshot].merge!(snapshot.wsi_snapshot || {})
+          state[:env_snapshot].merge!(snapshot.env_snapshot || {})
+          merge_env_dependency!(state[:env_dependency], snapshot.env_dependency || {})
+          merge_cache_hit_reason!(state[:cache_hit_reason], snapshot.cache_hit_reason || {})
+        end
+        # rubocop:enable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/PerceivedComplexity
+
+        # Per-example env attribution unions set-wise: an example that
+        # declared `tracks: { env: [A, B] }` on one worker and
+        # `tracks: { env: [B, C] }` on another (edge case; parallel_tests
+        # workers rarely run the same example) collapses to [A, B, C].
+        def self.merge_env_dependency!(target, source)
+          source.each do |id, names|
+            existing = target[id] || []
+            target[id] = (existing | Array(names)).sort
+          end
+        end
+
+        # Sum per-worker reason counts. parallel_tests partitions
+        # examples across workers; each worker's filtered_examples
+        # tally is disjoint by example_id, so sum is the right combine
+        # rule (a "Files changed" count from worker A plus the same
+        # reason's count from worker B = total examples that ran for
+        # that reason across the suite).
+        def self.merge_cache_hit_reason!(target, source)
+          source.each { |reason, count| target[reason] += count }
+        end
+
+        # Internal helper for the tracer pipeline.
+        # @api private
+        def self.merge_examples_coverage!(target, source)
+          source.each do |id, per_file|
+            entry = target[id] ||= {}
+            per_file.each do |file_path, lines|
+              file_entry = entry[file_path] ||= {}
+              lines.each do |line_key, strength|
+                file_entry[line_key] = (file_entry[line_key] || 0) + (strength || 0)
+              end
+            end
+          end
+        end
+
+        # Internal helper for the tracer pipeline.
+        # @api private
+        def self.reverse_of(dependency)
+          reverse = Hash.new { |h, k| h[k] = Set.new }
+          dependency.each do |id, file_names|
+            file_names.each { |name| reverse[name] << id }
+          end
+          reverse
+        end
+      end
+
+      private
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def last_run_path
+        File.join(@cache_path, LAST_RUN_FILENAME)
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def lock_path
+        File.join(@cache_path, LOCK_FILENAME)
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def maybe_prune_after_save
+        prune_run_dirs!(keep: @retention_local_count) if @retention_local_count
+      end
+
+      # Internal constant.
+      # @api private
+      BYTES_PER_MB = 1_048_576
+      private_constant :BYTES_PER_MB
+
+      # Emit a warn-level log line for each just-saved file that
+      # exceeded the per-file budget, and one total-budget line when
+      # the whole cache tree exceeds that threshold. Both thresholds
+      # are MiB; 0 / nil disables. The warning suggests the most
+      # effective remediations in order: add_filter for vendor paths
+      # (usually the biggest win), transitive_load_tracking off
+      # (cuts the constants-blind-spot overhead), and the `:msgpack`
+      # serializer (PR B). Budgets surface B11 symptoms (issue #15 /
+      # #20) without forcing behavior change.
+      def maybe_warn_size_budget(run_id)
+        warn_oversized_run_files(run_id) if positive_threshold?(@warn_per_file_mb)
+        warn_oversized_cache_total if positive_threshold?(@warn_total_mb)
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def positive_threshold?(value)
+        value.is_a?(::Integer) && value.positive?
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def warn_oversized_run_files(run_id)
+        run_dir = File.join(@cache_path, run_id)
+        return unless File.directory?(run_dir)
+
+        threshold_bytes = @warn_per_file_mb * BYTES_PER_MB
+        per_file_glob(run_dir).each do |path|
+          size = File.size(path)
+          next unless size > threshold_bytes
+
+          @logger&.warn(
+            "rspec-tracer cache: #{File.basename(path)} is #{format_mib(size)} " \
+            "(> #{@warn_per_file_mb} MiB per-file threshold); remediations (in order): " \
+            'add_filter for vendor paths, `transitive_load_tracking false`, ' \
+            '`storage_backend :json, serializer: :msgpack` for disk reduction'
+          )
+        end
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def warn_oversized_cache_total
+        total = total_cache_size_bytes
+        threshold_bytes = @warn_total_mb * BYTES_PER_MB
+        return unless total > threshold_bytes
+
+        @logger&.warn(
+          "rspec-tracer cache: total size is #{format_mib(total)} " \
+          "(> #{@warn_total_mb} MiB total threshold); remediations (in order): " \
+          '`cache_retention_local_count N` to cap history, ' \
+          'add_filter for vendor paths, ' \
+          '`storage_backend :json, serializer: :msgpack` for disk reduction'
+        )
+      end
+
+      # Glob matching this backend's serializer extension. Surfaces
+      # the active on-disk layout so size-budget warnings stay
+      # accurate when the user switches to `:msgpack` (.msgpack.gz
+      # files instead of .json).
+      def per_file_glob(run_dir)
+        Dir[File.join(run_dir, "*.#{@serializer.extension}")]
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def total_cache_size_bytes
+        total = 0
+        Dir[File.join(@cache_path, '**', "*.#{@serializer.extension}")].each do |path|
+          total += File.size(path) if File.file?(path)
+        end
+        total
+      rescue StandardError
+        0
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def format_mib(bytes)
+        "#{(bytes.to_f / BYTES_PER_MB).round(1)} MiB"
+      end
+
+      # Enumerate run-id subdirectories of cache_path, newest first
+      # by mtime. Non-directory children (last_run.json, lock file)
+      # and dotfiles are excluded so a stray `.DS_Store` or editor
+      # swap file doesn't confuse the prune math.
+      def collect_run_dirs
+        entries = Dir.children(@cache_path).filter_map do |name|
+          next if name.start_with?('.')
+
+          path = File.join(@cache_path, name)
+          next unless File.directory?(path)
+
+          [path, File.mtime(path).to_f]
+        end
+        entries.sort_by { |(_, mtime)| -mtime }.map(&:first)
+      end
+
+      # Split a newest-first list of run directories into (keep,
+      # prune). The live `current` run-id is always retained even if
+      # it fell off the top-N by mtime (defensive: an external
+      # `touch` on an old dir must not force deletion of the live
+      # one). Inputs are absolute paths; `current` is the basename
+      # reported by `last_run_id` (may be nil if last_run.json is
+      # missing or corrupt).
+      def partition_dirs_to_prune(candidates, keep:, current:)
+        keep_paths = []
+        prune_paths = []
+        candidates.each do |path|
+          if keep_paths.size < keep || File.basename(path) == current
+            keep_paths << path
+          else
+            prune_paths << path
+          end
+        end
+        [keep_paths, prune_paths]
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def read_last_run_manifest
+        return nil unless File.file?(last_run_path)
+
+        read_json(last_run_path)
+      rescue StandardError
+        nil
+      end
+
+      # last_run.json is always plain JSON regardless of serializer -
+      # it is the human-debuggable + CI-script-compatible pointer that
+      # USER_FACING_SURFACE.md section 6 locks. Helper stays narrow so the
+      # serializer dispatch can not accidentally reach it.
+      def read_json(path)
+        contents = File.read(path, encoding: ENCODING)
+        JSON.parse(contents)
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def write_json_atomic(path, data)
+        tmp_path = "#{path}.tmp.#{Process.pid}.#{rand(1_000_000)}"
+        File.write(tmp_path, JSON.pretty_generate(data), encoding: ENCODING)
+        File.rename(tmp_path, path)
+      ensure
+        File.delete(tmp_path) if tmp_path && File.file?(tmp_path)
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def write_run_files(dir, snapshot)
+        HASH_FIELDS.each { |f| write_run_field(dir, f, snapshot.send(f) || {}) }
+        ID_SET_FIELDS.each { |f| write_run_field(dir, f, serialize_id_set(snapshot.send(f))) }
+        DEPENDENCY_FIELDS.each { |f| write_run_field(dir, f, serialize_dependency(snapshot.send(f))) }
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def write_run_field(dir, name, payload)
+        write_payload_atomic(File.join(dir, field_filename(name.to_sym)), payload)
+      end
+
+      # Atomic write for a per-field payload. Encodes via the active
+      # serializer; writes binary so msgpack + zlib bytes are not
+      # re-encoded by Ruby's IO layer. Same tmp-rename pattern as
+      # write_json_atomic (last_run.json commit point).
+      def write_payload_atomic(path, data)
+        tmp_path = "#{path}.tmp.#{Process.pid}.#{rand(1_000_000)}"
+        File.binwrite(tmp_path, @serializer.encode(data))
+        File.rename(tmp_path, path)
+      ensure
+        File.delete(tmp_path) if tmp_path && File.file?(tmp_path)
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def write_last_run_atomic(schema_version:, run_id:)
+        manifest = { 'schema_version' => schema_version, 'run_id' => run_id, 'timestamp' => Time.now.utc.iso8601 }
+        write_json_atomic(last_run_path, manifest)
+      end
+
+      # Dispatch one field's raw JSON body through the right
+      # deserializer. Paired with FIELD_KINDS. Kept here rather
+      # than on the reader so all shape knowledge stays in one
+      # class; mutant sees one AST node per kind branch.
+      def decode_field(field, raw)
+        case FIELD_KINDS.fetch(field)
+        when :symbolized    then deserialize_symbolized(raw)
+        when :dupe_examples then deserialize_dupe_examples(raw)
+        when :id_set        then deserialize_id_set(raw)
+        when :dependency    then deserialize_dependency(raw)
+        when :plain_hash    then deserialize_plain_hash(raw)
+        end
+      end
+
+      # Plain-hash round-trip: JSON.parse returns nil on failure, and
+      # a valid-but-wrong-shape file would otherwise poison the
+      # Snapshot field. Treat anything non-Hash as the empty default.
+      def deserialize_plain_hash(raw)
+        raw.is_a?(Hash) ? raw : {}
+      end
+
+      # Walk a parsed JSON tree and replace every String with its
+      # frozen-string-table entry via `String#-@`. Idempotent on
+      # already-frozen Strings. Portable across every matrix Ruby
+      # (json-gem's `freeze: true` option only arrived in 2.8 /
+      # Ruby 3.4); the explicit walk keeps behavior identical on
+      # Ruby 3.1 + 3.2 cells.
+      #
+      # Big win on dependency.json where the same file path repeats
+      # across every example that depends on it - 2000 unique paths
+      # may appear 1M+ times; interning collapses that to 2000
+      # objects + refs. Small overhead on fields where strings are
+      # unique (description text in all_examples) but the RAM
+      # savings on the path-heavy fields dominate.
+      def deep_intern(obj)
+        case obj
+        when Hash
+          obj.each_with_object({}) { |(k, v), h| h[k.is_a?(String) ? -k : k] = deep_intern(v) }
+        when Array
+          obj.map { |v| deep_intern(v) }
+        when String
+          -obj
+        else
+          obj
+        end
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def read_run_file(dir, name)
+        path = File.join(dir, name)
+        return nil unless File.file?(path)
+
+        @serializer.decode(File.binread(path))
+      rescue StandardError
+        nil
+      end
+
+      # Sorted Array on disk, Set in memory - matches 1.x report_writer.
+      def serialize_id_set(collection)
+        return [] if collection.nil?
+
+        collection.to_a.sort
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def deserialize_id_set(raw)
+        return Set.new unless raw.is_a?(Array)
+
+        raw.to_set
+      end
+
+      # dependency / reverse_dependency: Hash[id => Set<path>] in
+      # memory, Hash[id => Array<path>] on disk.
+      def serialize_dependency(collection)
+        return {} if collection.nil?
+
+        collection.transform_values { |paths| Array(paths) }
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def deserialize_dependency(raw)
+        return {} unless raw.is_a?(Hash)
+
+        raw.transform_values { |paths| Array(paths).to_set }
+      end
+
+      # all_examples / all_files: inner hashes whose keys 1.x
+      # symbolizes after parse.
+      def deserialize_symbolized(raw)
+        return {} unless raw.is_a?(Hash)
+
+        raw.transform_values do |inner|
+          inner.is_a?(Hash) ? inner.transform_keys(&:to_sym) : inner
+        end
+      end
+
+      # duplicate_examples: Hash[id => Array<Hash>]; each inner hash's
+      # keys symbolized post-parse.
+      def deserialize_dupe_examples(raw)
+        return {} unless raw.is_a?(Hash)
+
+        raw.transform_values do |list|
+          next list unless list.is_a?(Array)
+
+          list.map { |entry| entry.is_a?(Hash) ? entry.transform_keys(&:to_sym) : entry }
+        end
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def info(message)
+        @logger&.info(message)
+      end
+
+      # Map the user-facing `:json` / `:msgpack` name onto the
+      # concrete Serializer class. A `:msgpack` request with the
+      # msgpack gem absent warns once and falls back to `:json` so
+      # the user's test suite keeps running (graceful-degradation
+      # contract; same posture as the remote-cache optional deps).
+      def resolve_serializer(name)
+        case name
+        when :json
+          Serializer::Json
+        when :msgpack
+          Serializer::Msgpack.available? ? Serializer::Msgpack : msgpack_unavailable_fallback
+        else
+          raise ArgumentError,
+                "unknown serializer: #{name.inspect}; allowed: [:json, :msgpack]"
+        end
+      end
+
+      # Internal method on the tracer pipeline.
+      # @api private
+      def msgpack_unavailable_fallback
+        @logger&.warn(
+          'rspec-tracer cache: msgpack gem is not installed; falling back to :json. ' \
+          "Add `gem 'msgpack'` to your Gemfile to use the :msgpack serializer."
+        )
+        Serializer::Json
+      end
+    end
+    # rubocop:enable Metrics/ClassLength
+  end
+end