Release v0.4.0 "vegan leather"

[TGPSKI]

Release v0.4.0 "vegan leather" — cut from release-v0.4.0, targeting main.

This is a reliability-focused release hardening the tannery curing pipeline under real concurrency (found load-testing examples/11-high-volume-ci at 100-webhook bursts), plus reasoning-model support and a new shell-completion command.

Highlights

Added

• leather completion <shell> — static bash/zsh/fish completion scripts mirroring the dispatch table.
• Reasoning-aware completion_reserve — per-agent override plus a new reasoning_reserve split so <think> traces and answers get separate budgets; model-aware defaults for Qwen3/QwQ/DeepSeek-R1 and a self-healing retry on length-truncation.
• thinking: agent front-matter — thinking: false disables a model's hidden reasoning trace per agent (measured 5.2x speedup on the 40-webhook burst with equal-or-better correctness).
• shell-mcp per-argument patterns validation and a full-system load-test profiler (scripts/profile/).

Fixed

• Concurrency correctness in the curing worker: shared-agent prompt corruption, lost work on transient failure, fan-in group retry/DLQ, and mislabeled artifact hide_kind (fix: curing collect fan-in hangs forever and leaks state when a leg permanently DLQs #45).
• Collect fan-in (collect_size) no longer hangs forever / leaks state when a leg permanently DLQs — new collect_timeout_seconds (curing: collect fan-in (collect_size) hangs forever and leaks state when one leg permanently DLQs #44).
• Curing-driven agents now inherit the global default model; prefix-scan curings honor configured concurrency; shell-tools.json schemas emit required (fix: curing model resolution, prefix-curing concurrency, and shell-tools.json schemas #43).
• System-prompt-only agents accepted by strict backends (runner: system-prompt-only agents rejected by strict OpenAI-compatible backends ("No user query found in messages") #41); HTTP client timeout no longer races the context deadline; 32-bit ID suffixes end burst-load collisions; reasoning-model tool-call loop and dropped-answer-text flakiness addressed (fix: reasoning-model flakiness, thinking-mode support, and high-volume-ci tuning #46).

Removed

• leather chat (no deprecation cycle — use leather run).
• ROADMAP.md and its references — deferred-item tracking moved to GitHub issues.

Release-prep housekeeping

• CHANGELOG.md: cut the [0.4.0] section and updated compare links.
• SECURITY.md: Supported Versions now lists 0.4.x; 0.3.x retired; trust-model heading bumped to v0.4.

See CHANGELOG.md for the full entry.