Feat: Pre-fill derive topics from the scope at point; add learning tests to the TDD menu

[tninja]

Two follow-ups to #517 that make topic-based derivation quicker to reach from code.

Pre-fill the topic from the scope at point. Every derive command in ai-code-doc.el starts by asking for a topic, and when you invoke one from a source file the topic you want is usually the function or class you are standing in, so typing it by hand is busywork. The new ai-code--topic-at-point reuses ai-code--current-qualified-scope-name (Tree-sitter first, then which-function) and, in prog-mode buffers, returns something like Service.run (src/Service.java), with the path relative to the Git root (just the scope name when the buffer visits no file). The two topic readers use it differently:

• Document topics (ai-code--read-document-topic): an empty topic there means "whole repository", so silently pre-filling would change the default. It first asks Scope document to "Service.run (src/Service.java)"? — n keeps the whole-repo document without reading a topic, y pre-fills the topic for editing.
• Unit test topics (ai-code--read-unit-test-topic): the topic is required anyway, so it is pre-filled directly with no extra question.

Outside prog-mode, or when no scope is found, both prompts behave exactly as before.

Learning tests in C-c a t. ai-code-derive-topic-unit-tests was only reachable from the architecture menu (C-c a A), but test-related actions are naturally looked for in the TDD menu. ai-code-tdd-cycle now ends with 6. Learning tests (Derive unit tests as a tutorial for a topic), which dispatches to the same command; the C-c a A entry stays. In Magit buffers C-c a t keeps its existing hunk-based path, so the option is not offered there.

Verification.

• New tests, each written first as a failing test: three in test/test_ai-code-doc.el (accepting and declining the scope question, no offer outside prog-mode, unit test topic pre-fill) and one in test/test_ai-code-agile.el (last TDD choice and its dispatch).
• Full suite: emacs -batch -L . --eval '(setq package-user-dir (expand-file-name "no-such-elpa" temporary-file-directory) native-comp-enable-subr-trampolines nil)' -l ert --eval "(mapc #'load-file (file-expand-wildcards \"test/test_*.el\"))" -f ert-run-tests-batch-and-exit → 1586 tests, 0 unexpected, 13 skipped. Disabling subr trampolines is only a local workaround for native-comp errors when cl-letf stubs completing-read / read-string.
• Byte-compile and checkdoc on the touched files add no new warnings (the helper uses when-let* to avoid the when-let obsolescence warning in newer Emacs).

Known caveat, left alone on purpose: with behavior auto-presets enabled, ai-code-tdd-cycle maps to the tdd-dev preset, so a learning-test prompt started from C-c a t also picks up +tdd. README.org is not updated for either change yet.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5682e93c12

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Normalize the file before computing the topic path

When a repository is opened through a directory symlink, ai-code--git-root returns the canonical file-truename, while buffer-file-name can retain the symlinked path. Relativizing those two different path namespaces produces a value such as ../../repo-alias/src/Service.java instead of src/Service.java, so accepting the pre-filled topic sends the AI a misleading path outside the reported repository root. Normalize the visited file with buffer-file-truename or file-truename before calling file-relative-name.

Useful? React with 👍 / 👎.