kurvan not in kosha data

[avinashvarna]

Unless I am doing something very wrong, kurvan and all shatranta forms of kf are not in the kosha data v0.4.0.
I tried narrowing it down within https://github.com/ambuda-org/vidyut/blob/main/vidyut-data/src/bin/create_kosha.rs, but couldn't find the root cause due to my lack of familiarity with rust.
If I just try creating the kosha with the single dhatu, kf, I see that the entry for kurva is present at the end of line

vidyut/vidyut-data/src/bin/create_kosha.rs

Line 916 in 487836b

}

but is not in the final kosha.

Is this a real bug, or am I missing something?

[avinashvarna]

I discovered that form in kosha vs kosha[form] for form = "kurva" / "kurvan" were returning opposite results.

After debugging with help from Gemini, this is the issue apparently:

1. Missing Final Output in Prefix Identification: The contains_suffix method was identifying prefixes by calling to_packed_entry(out_base). However, the out_base (the output accumulated along edges) does not include the final_output of the current node. If a node represents both a complete word (e.g., kurva as a verb) and a prefix for another word (e.g., kurva as a prefix for kurvan), the out_base alone might map to the verb's POS, while the out_base + final_output maps to the SubantaPrefix.
2. Incorrect contains_key Logic: contains_key was returning node.is_final(). This is insufficient because a node can be final while only representing a SubantaPrefix or TinantaPrefix. A key is only truly "contained" in the kosha if it represents a complete word (Tinanta or Subanta).

I did verify that the changes it suggests do fix the issue. However, I am not sure what the policy on AI generated code is. In this case, it is minimal - just a couple of lines changed to fix these issues. If that is OK, I can submit a PR to fix the issue. Otherwise, hopefully this info is sufficient for someone else to fix the issue.

IMHO, this is a serious enough bug in the kosha logic that it warrants a new release once fixed.

[akprasad]

yes, please make a PR! thanks for debugging this.