refactor(llm-agents): replace flattening overlay with upstream shared-nixpkgs

Drop the custom mkLlmAgentsOverlay factory (172 top-level attrs +
17-entry blocklist guarding against nixpkgs clobbering — the source of
the webcord buildNpmPackage and fetchPnpmDeps stack overflows).

Apply inputs.llm-agents.overlays.shared-nixpkgs directly instead. It
exposes everything under the pkgs.llm-agents.* namespace and clobbers
nothing, so the blocklist becomes unnecessary. Packages stay built
against this system's nixpkgs (allowUnfree), which qmd.override
cudaSupport in the m3-ares NVIDIA specialisation requires.

Consumers switched to namespaced refs:
- hosts/m3-hermes/services/hermes-agent.nix: herdr, opencode, qmd
- m3ta-home agents.nix (input bumped): agent-browser, beads, herdr,
  qmd, openspec — unqualified refs would now silently resolve to
  outdated nixpkgs attrs (opencode 0.3.x) or vanish.

openspec pass-through dropped from m3ta-nixpkgs (input bumped):
llm-agents is the single source of truth now.

Note: flake.nix/lock carry the pre-existing local path inputs for
m3ta-home and m3ta-nixpkgs (dev state, gitea URLs kept as comments).

Verified: drvPath evals for m3-hermes, m3-atlas, and the m3-ares
NVIDIA specialisation (qmd + CUDA EULA) all pass; statix/deadnix clean.

hosts/common/default.nix

@@ -40,7 +40,11 @@
 
       inputs.m3ta-nixpkgs.overlays.default
       # inputs.m3ta-nixpkgs.overlays.modifications
-      (outputs.lib.mkLlmAgentsOverlay system)
+      # llm-agent packages under the `pkgs.llm-agents.*` namespace, built
+      # against this system's nixpkgs (needed for qmd.override cudaSupport
+      # with allowUnfree). Do NOT flatten to top-level attrs — see
+      # overlays/default.nix.
+      inputs.llm-agents.overlays.shared-nixpkgs
       # You can also add overlays exported from other flakes:
       # neovim-nightly-overlay.overlays.default