refactor: standardize environment settings and generate their reference

[EItanya]

Kagent-specific environment variables had inconsistent names, and the registry was missing settings used by the controller, CLI, runtimes, Python packages, and UI. This PR standardizes those names under KAGENT_ and generates docs/env.md from the existing ENV registry. The reference covers 138 user-configurable settings; controller-owned payloads, generated credentials, and private runtime wiring are excluded.

This is a direct pre-release rename. Custom deployment environments and local configuration need to use the new names, for example LOG_LEVEL → KAGENT_LOG_LEVEL, POSTGRES_DATABASE_URL → KAGENT_POSTGRES_DATABASE_URL, and the Go ADK's PORT → KAGENT_PORT. The old names are no longer read by kagent. Helm output, runtime injection, scripts, and examples use the new names. Upstream SDK and tool conventions such as OTEL_*, provider credentials, and KUBECONFIG stay unchanged.

• make env-docs generates the reference through kagent env. make env-docs-check checks freshness without modifying it and runs in the existing manifests-check CI job. The workflow includes changes to docs/env.md while skipping other Markdown-only changes, with no additional job. Generation exports registered metadata, never values from the current process environment.

• Registration functions accept multiple components so shared settings only need one registration. Filters match any listed component; JSON exports a components array instead of the previous component string.

• Go controller, CLI, ADK, provider, logging, telemetry, E2E, and upgrade-test settings read through registered accessors. Telemetry rendering takes its environment names and protocol default from the same registry. Get() and Lookup() preserve component-specific defaults and precedence. Boolean/integer LookupWithError() preserves diagnostics for invalid migration and capture-limit settings. Go booleans accept true/t/1 and false/f/0 regardless of case or surrounding whitespace; upstream OTEL_SDK_DISABLED retains its true-only grammar. The Python ADK HTTP server uses shared KAGENT_LOG_LEVEL; the separate UVICORN_LOG_LEVEL override is removed.

• The existing Go linter rejects direct os.Getenv/os.LookupEnv reads, including aliased imports and function references. E2E and upgrade tests are covered by the rule; unit tests may inspect process state directly. The registry itself and explained exceptions for generated payloads, credential placeholders, and the standalone BYO example remain allowed.

• Both ADKs use KAGENT_PORT for A2A traffic. The controller supplies 80 once; Python's KAGENT_A2A_GRPC_ADDRESS is removed. Standalone defaults and explicit programmatic/CLI overrides are preserved, and Python's HTTP listener remains separately configured.

• Chart-defined non-secret settings live in controller, UI, and bundled PostgreSQL ConfigMaps, loaded through envFrom. Pod-template checksums trigger rollouts when configuration changes. Credentials, downward-API values, values needing Kubernetes $(...) expansion, and explicit user env overrides stay in the Deployments.

• Testing settings use KAGENT_TEST_* or KAGENT_E2E_*; UI settings use KAGENT_UI_*. UI containers and Vite consume the same names, including KAGENT_UI_ENABLE_MOCK. Browser configuration is limited to explicit UI settings, KAGENT_UI_EXTENSION_*, and build-time KAGENT_UI_VITE_* switches. Make targets, CI, Playwright, and local-development examples use the same prefixes. Upgrade tests also register their repository, version, registry, cluster, namespace, and context inputs under KAGENT_E2E_*; Make maps its existing build variables to those names. Existing empty-value fallbacks and true-only opt-in grammars are preserved.

• Seven settings ignored by the controller are removed along with their Helm values and references. They covered authentication, gRPC TLS/message limits, and the default Substrate worker pool. The chart's MCP URL uses HTTP to match the controller listener. [BUG] Restore Helm-configured controller authentication end to end #2968 tracks restoring Helm-configured authentication end to end.

• Image-digest lookups in CI and cluster setup consume the full Docker output. Exiting awk early can close the pipe while Docker is still writing and abort E2E setup under pipefail.

Validation passed:

• Focused Go tests for the registry, controller app, CLI/database commands, ADK app/A2A/agent/providers/embedding/tools, logging, telemetry, harness configuration, and translators; make -C go lint (zero issues). Coverage checks unset values, invalid input, capture limits, and boolean grammars. Negative lint probes verify that aliased direct reads and function references are rejected, including in E2E test files. Focused configuration tests cover E2E URL precedence, mock-host overrides, and upgrade defaults/context selection; Make dry runs verify that both upgrade targets pass the registered names.

• go build -mod=readonly passes for the Go ADK, both harnesses, and sandbox guest.

• Sandbox guest Docker builder passes for linux/amd64 and linux/arm64 with -mod=readonly. Its minimal source context includes the ENV package required by logging.

• make env-docs-check, including checks that stale or missing docs fail and a failed generator preserves existing documentation.

• Helm unit tests (369 tests) and chart lint. Before/after renders preserve effective container environments across default, custom, external-database, and database-file configurations; setting changes update rollout checksums.

• Python CLI, A2A configuration, and gRPC tests, plus Ruff on changed Python files. A real gRPC connection verifies the shared port is consumed and explicit addresses take precedence.

• UI typecheck, focused unit tests (42 tests), and lint (no errors). Vite and container-startup smoke checks verify prefixed settings reach the browser and unrelated environment values stay private. Playwright configuration checks verify the renamed ports and mock/live switches; Upgrade packages compile without running cluster tests.

• Full E2E suite passed on a clean local Kind cluster using images built from this branch: 52 top-level tests, zero failures, and 10 existing subtest skips in 7m01s, with four concurrent scenarios and tracing required. This covers all four harnesses, CLI invocation, checkpoints, controller restarts, scheduled runs, runtime revision lifecycle, and sandbox workflows. The BYO fixture uses KAGENT_PORT=80; the old PORT=80 left the runtime listening on 8080 while the gateway connected to 80. The interaction test reproduced that timeout before the fixture change and passed afterward.