Summary

Closes the residual hidden-input gap noted in PR #52: when TUP_PLATFORM (or any env var) is sourced from the environment and used inside a conditional like ifeq ($(TUP_PLATFORM),win32), flipping it between builds did not trigger a rebuild — the output stayed stale and the build reported "Nothing to do."

The root cause: an asymmetry the builder built up over time

The builder already folded config vars used in a guard into every guarded command's identity, via condition_config_vars → Sticky edge. The env-var path had no equivalent:

When env TUP_PLATFORM flipped between builds, the phi-node's two branch commands sat in the graph unchanged (same text, same outputs, same identity), the output file on disk was unchanged, and change detection saw nothing to do.

What changed

Symmetric mirror of the config machinery, end to end:

• condition_env_vars added to BuilderContext (include/pup/graph/builder.hpp), symmetric to condition_config_vars.
• process_conditional captures used_env_vars from the condition's expansion and merges it into condition_env_vars under the same ScopeGuard that already handles nested conditionals.
• create_command_node emits Sticky edges from those env Variable nodes to every command in the block — exact mirror of the existing config loop.

And to make sure those sticky edges land on real nodes:

• ensure_env_var_node(ctx, state, name, value) extracted from process_import (its Variable-node creation/update block). process_import is refactored to delegate to it, with no behavioral change to its env → cached → default fallback.
• on_env_var_used callback now calls the helper too, so TUP_PLATFORM/TUP_ARCH — auto-resolved from getenv without an explicit import — get the same tracking node any imported var would.

Effect on command identity

compute_command_identity (from PR #52) already folds every Sticky-linked Variable node, so no change is needed in dag.cpp. And no on-disk format change: the new sticky edges are persisted by serialize_edges automatically.

Test plan

• Reproducer [platform-incremental] + fixture platform_conditional_incremental: both ifeq branches produce the same output file with distinct content; the platform string never appears in either command's text, so only the sticky edge can detect the flip. Build with TUP_PLATFORM=win32 (asserts WINBUILD, no-op stability), then re-bind to linux (asserts !is_noop() and NIXBUILD). Red before this change, green after.
• Full suite green: 466 cases / 23,671 assertions.
• Targeted regression coverage: [config] (the symmetric config path stays intact), [import] (env_var_persist, env_dep_subprocess, equals-in-value), [platform] (first-build platform_conditional), [idshift] and [envdep] (the regression guards from Identify commands by a structural hash, not a string or position #52).
• clang-format, clang-include-cleaner, and scoped clang-tidy on changed files are clean for new code (the file's pre-existing std::move-on-trivially-copyable / qualified-auto style is unchanged).

Out of scope

• TUP_PLATFORM/TUP_ARCH resolved from the compile-time default (Builtin bank) need no tracking — constant per putup binary.
• No graph-wide refactor of implicit-dep representation; targeted at the env-var → condition asymmetry only.

🤖 Generated with Claude Code