fix: robust task-CLI I/O (#5), prefix-preserving completion (#4), wrap option + focusable confirm picker (#3)

[MattHandzel]

What this fixes

#5 — :Tw opens a blank buffer, :TwRefresh has no effect

Root cause. vim.fn.system merges stderr into stdout. Any Taskwarrior chatter around the export JSON — the TW 3.x task news nag, Configuration override notices (triggered by our own rc.* args for users with verbose=override), hook output, forced-color ANSI escapes — made the JSON parse fail, and every parse site silently degraded to "zero tasks". A zero-task render is only the header comment, which conceallevel=3 hides entirely: the buffer looks blank, and :TwRefresh faithfully re-renders the same empty result. (Telling detail from the report itself: the feedback template's task_count_bucket worked — 11-25 — because that one call uses 2>/dev/null.)

Fix, defense in depth:

1. READ_RC adds rc.verbose=nothing rc.color=off to every read-only task invocation — chatter suppressed at the source. Mutation paths keep verbose output (tw_add parses Created task <uuid>).
2. New decode_json_array: strips ANSI, tries prioritized bracket candidates, and falls back to line-filtering for chatter interleaved between TW3's line-per-task output.
3. The ten hand-rolled find("%[") slices (buffer render/virt-text, graph, inbox, views, delegate ×2, modify, statusline, telescope picker, api.export) are replaced by one shared taskmd.shell_export that additionally redirects stderr away from the parse stream.
4. tw_export now raises a diagnosable error (with the actual task output) instead of silently returning {} — a failure can no longer render as silence.
5. A genuinely-zero-task render shows a virt-lines empty-state hint (No tasks match filter: … · :TwFilter to change it) instead of a blank buffer.
6. _udas / _projects / _tags line parsing rejects lines containing whitespace (valid names never do; chatter always does) so nag lines can't become phantom UDAs or completion candidates.

#4 — Tab in the :TwFilter box deleted the already-typed filter

vim.fn.input() completion replaces the entire typed text with the chosen candidate; _complete_filter returned bare last-token candidates, so completing +ACTIVE after project:foo wiped project:foo. Candidates now carry the untouched leading tokens. Bonus find: the gm modify prompt referenced _complete_modify, which was never defined — Tab there raised E117. It's now implemented, and a lint spec asserts every v:lua completion reference in the codebase resolves to a real function.

#3 — wrap option + unusable Apply/Cancel popup

• New setup{ wrap = false } option (validated, documented in README + :help) for one-line-per-task rendering.
• The Apply/Cancel confirmation was a vim.ui.select called synchronously inside the BufWriteCmd handler; float-based select backends (dressing, snacks, telescope) rendered the picker but couldn't take focus — cursor stayed in the task buffer. The picker is now vim.schedule()d out of the autocmd context.

Verification

• 19 new unit specs (tests/lua/spec/issue_fixes_spec.lua): decode robustness (leading/trailing/bracketed/ANSI/interleaved noise), tw_export raising instead of silently returning {}, empty-state extmarks, completion prefix preservation, _complete_modify, the v:lua lint, wrap validation.
• 6 new e2e specs against a real Taskwarrior DB: render survival under a real noisy on-launch hook, wrap window state, empty-state extmark on a real render + absence when tasks match, and the confirm-picker Apply/Cancel drain.
• Full Lua unit suite and e2e suite green (the one pre-existing icons=true → NF glyph e2e failure exists on main and is unrelated — worth a separate issue).

Closes #3. Closes #4. Addresses #5 — leaving that one to be closed once @oliviertripet confirms on this branch (repro depends on his environment; see the ask in the issue thread).

🤖 Generated with Claude Code