feat(export): persistent on-disk helmTemplate cache

[julienduchesne]

Summary

Exporting many environments that render the same Helm charts pays the helm-rendering cost on every run. Add an experimental, opt-in --helm-cache flag to rtk export that keeps a helm-cache/ directory inside the output dir and reuses helmTemplate results across runs and environments.

• Cache hits skip both the helm subprocess and YAML parsing: the cached value is the manifested JSON of the rendered resource map, so a hit is a single deserialize into a Val
• Cache key now covers the full call — release name, chart path, namespace, values, flags, nameFormat, and the chart's Chart.yaml contents (so a vendored version bump invalidates stale entries)
• Add RTK_HELM_DISABLE_MEMOIZATION to bypass the in-memory cache entirely (helm runs on every call) for benchmarking

Context

Builds on the process-global in-memory Helm cache. That cache now stores the post-parse JSON projection instead of raw YAML, so hits no longer re-parse YAML, and it is shared across worker threads — a chart rendered for one environment is reused by every other. The on-disk cache is global and loaded/written exactly once per run: before the parallel loop every entry is preloaded into memory and the directory is cleared; after the loop only the keys touched this run are written back, which prunes stale entries. A process-global touched set is used because a single main.jsonnet can expand into many environments across worker threads (an earlier per-environment design raced on the shared directory).

Gated behind --helm-cache (off by default) so the metadata directory never appears in normal exports or golden-test output.

Stacked on #3.

Made with Cursor

The base branch was changed.

[github-actions]

Benchmark Results

Benchmark	Test	vs tk	vs base
Diff	cluster_scoped	48.73x faster	equal
Diff	configmap_unchanged	30.87x faster	equal
Diff	deployment_nested_changes	28.42x faster	equal
Env List	Single Inline Directory (--json)	46.32x faster	equal
Env List	Single Inline File (--json)	46.09x faster	equal
Env List	All Environments (--json)	84.41x faster	equal
Eval	Single Static Environment	15.76x faster	equal
Eval	Inline Environment File	19.02x faster	equal
Eval	Eval Expression - Deeply Nested Path	15.86x faster	equal
Export (Full)	Single Static Environment	6.96x faster	equal
Export (Full)	Single Inline Environment	21.27x faster	equal
Export (Full)	All Environments	19.25x faster	1.02x slower
Export (Replace)	Single Static Environment	5.69x faster	equal
Export (Replace)	Single Inline Environment	21.61x faster	equal
Export (Replace)	All Environments	27.94x faster	1.01x faster
Helm Template	Full Export	1.41x faster	equal
Tool Importers Count	Single Lib Directory (Non-Recursive)	2.24x faster	equal
Tool Importers Count	Single Lib Directory (Recursive)	2.23x faster	equal
Tool Importers Count	Entire Lib Directory (Non-Recursive)	1.34x faster	equal
Tool Importers Count	Entire Lib Directory (Recursive)	1.14x faster	equal
Tool Importers	Global Lib File	3.29x faster	equal
Tool Importers	Single Env-Specific Lib File	1.58x faster	equal
Tool Importers	All Env-Specific Lib Files	5.2x faster	equal
Tool Importers	All Lib Files	5.23x faster	equal
Tool Importers	All Jsonnet Files	9.04x faster	equal
Tool Imports	Single Static Environment	3.38x faster	equal
Tool Imports	Single Inline Environment File	3.48x faster	equal
Validate Manifests	All Validations	-	equal
Validate Manifests	Manifest Tests Only	-	equal
Validate Manifests	Namespace Test Only	-	equal

Full results available in workflow artifacts.

Benchmark run on commit afc832220e78e518d527ced7ac53120a49cb0d78