vm: coerce the path builtin's path argument like Nix's coerceToPath

[dan-v]

Problem

builtins.path rejects an attrset path argument with a type error. Nix's prim_path routes it through coerceToPath: paths and strings directly, __toString recursively, anything else through the non-copying string coercion (an attrset via outPath), then requires an absolute result. Passing a flake input or derivation-like value as path is a common idiom.

$ nix-instantiate --eval --impure -E 'builtins.path { path = { outPath = ./tree; }; name = "t"; }'
"/nix/store/<hash>-t"
$ fix eval --impure -E 'builtins.path { path = { outPath = ./tree; }; name = "t"; }'
error: type error

Fix

Route the path attribute through the existing non-copying string coercion (stringLikeValue, whose __toString-before-outPath ordering matches coerceToPath). The absolute-path requirement and the other attributes are unchanged.

Verification

• outPath-as-path, outPath-as-string, __toString, and a non-canonical /x/../. argument all produce the store path Nix produces, byte-identical.
• A unit test pins the four vectors against the plain-string argument.
• Lix (357) and snix (114) conformance corpora pass; a nixpkgs universe chunk of 9,732 attrs has zero drvPath mismatches.

[psyclyx]

Merged as c5913653, preserving the PR's original author attribution. Verified against nix-instantiate: the outPath-as-path, outPath-as-string, and __toString arguments all produce the identical store path, and __toString correctly wins over outPath.

Three follow-up commits on top. stringBytes only borrows, and for a heap string the slice dies at the next GC safepoint, so routing the path attribute through the string coercion (a __toString result is a fresh heap string) left it dangling across the later attr forces and the user filter calls in recursiveIngest; the coerced bytes are now interned, which covers store_name too. The name and sha256 attrs had the same defect already and got the same treatment.

The third one is the crash you would have hit next: a cyclic { outPath = r; } reaching the coercion recursed on the native stack until the fiber faulted. Nested coercion levels now count against max-call-depth, as Nix's coerceToString does via addCallDepth, with the same native backstop forceThunkImpl uses. The depth is fiber state beside tryeval_depth, so every clone of the walk shares one budget. Cyclic outPath, __toString, nested-list, and builtins.path shapes all report stack overflow; max-call-depth exceeded now instead of segfaulting, and finite chains are unchanged.

The derivation-specific coercion clones still recurse unguarded; #8 covers those and can adopt coercionEnter/coercionExit. Thank you!