Add prompt editing support for task files

[tninja]

Summary

Task files are becoming the main place I write prompts, so this adds editing support aimed at that workflow: @ reference gating, outline navigation, long-form writing helpers, and reuse of prompts already written.

The immediate motivation was that @ completion fired on every @ character. Prompt files routinely contain pasted Java or Python, where @ is an annotation, not a file reference, so pasting code triggered a blocking completion prompt repeatedly. A single guard, ai-code--prompt-reference-position-p, now decides: @ must be outside any literal Org block and must begin a word. Both the capf path and the post-self-insert-hook path call that one guard, so they cannot drift apart. The explicit C-c @ command deliberately bypasses it, so completion is still reachable inside a src block when you actually want it.

Changes

• Reference gating — mode-agnostic literal-block detection (textual, not org-in-src-block-p, so it also holds in non-org buffers); rejects @Override and user@, accepts @ at word start. Opts the mode out of helm's completion-in-region via helm's own supported option rather than advising helm internals, and enables company-capf buffer-locally so global-company-mode stays off.
• Navigation — jump to a heading by full outline path, focus/restore a subtree, and per-file startup folding. New task files get #+STARTUP: content; folding applies only to files under the task directory, so the global org-startup-folded setting is untouched.
• Writing — flyspell skips literal blocks (pasted code and logs are not prose), C-c b wraps a region in a block, and an opt-in ai-code-prompt-writing-mode for visual wrapping. Display only; buffer text is never rewritten, so what gets sent to the agent is unchanged.
• Reuse — search past task files as a prompt corpus, browse the existing snippets by descriptive name instead of terse key, plus small context helpers for file/branch/defun.

All new keybindings are bound at load time rather than inside the mode body, so describe-mode lists them before the mode has been activated. Every soft dependency (helm, company, visual-fill-column, yasnippet) degrades silently when absent, with tests on the missing-dependency path.

Verification

A gauntlet script (test/run_ai_code_prompt_editing_gauntlet.sh) runs eight layers as fail-closed gates: 62 focused tests green, full suite at 1482 tests with the 13 pre-existing failures unchanged (names verified identical to baseline), byte compilation clean, checkdoc clean, 6/6 mutants killed by failing assertions, a negative control proving the guard tests actually bite, and real execution against a live task file in a real Emacs.

Two findings worth flagging, both fixed: a void-variable on helm's option aborted mode activation entirely and was caught only by the real-execution layer — the unit tests had been masking it — and one mutant initially "killed" the suite by making the file unloadable rather than by being detected, so the gauntlet now treats an incomplete run as a failure instead of evidence.

Scope

Additive. No change to the backend send path, no new hard dependencies, and the existing snippets are untouched. The pre-existing version< bug in test/test_00-bootstrap.el is out of scope; the gauntlet works around it with an isolated package-user-dir.